feat(backend): implement unified content embeddings with userId support

- Replace StoreListingEmbedding with UnifiedContentEmbedding table
- Add ContentType enum (STORE_AGENT, BLOCK, INTEGRATION, DOCUMENTATION, LIBRARY_AGENT)
- Support user-specific content with optional userId field for access control
- Maintain backward compatibility with wrapper functions for existing store APIs
- Update hybrid search to use unified embedding table with proper ContentType filtering
- Add comprehensive tests for new embedding service functionality
- Use proper Prisma ContentType enum instead of strings for type safety

The unified architecture enables future expansion to semantic search for blocks,
documentation, and library agents while maintaining existing store functionality.

autogpt_platform/backend/backend/api/features/store/backfill_embeddings.py

@@ -0,0 +1,72 @@
+#!/usr/bin/env python3
+"""
+CLI script to backfill embeddings for store agents.
+
+Usage:
+    poetry run python -m backend.server.v2.store.backfill_embeddings [--batch-size N]
+"""
+
+import argparse
+import asyncio
+import sys
+
+import prisma
+
+from backend.api.features.store.embeddings import (
+    backfill_missing_embeddings,
+    get_embedding_stats,
+)
+
+
+async def main(batch_size: int = 100) -> int:
+    """Run the backfill process."""
+    # Initialize Prisma client
+    client = prisma.Prisma()
+    await client.connect()
+    prisma.register(client)
+
+    try:
+        # Get current stats
+        print("Current embedding stats:")
+        stats = await get_embedding_stats()
+        print(f"  Total approved: {stats['total_approved']}")
+        print(f"  With embeddings: {stats['with_embeddings']}")
+        print(f"  Without embeddings: {stats['without_embeddings']}")
+        print(f"  Coverage: {stats['coverage_percent']}%")
+
+        if stats["without_embeddings"] == 0:
+            print("\nAll agents already have embeddings. Nothing to do.")
+            return 0
+
+        # Run backfill
+        print(f"\nBackfilling up to {batch_size} embeddings...")
+        result = await backfill_missing_embeddings(batch_size=batch_size)
+        print(f"  Processed: {result['processed']}")
+        print(f"  Success: {result['success']}")
+        print(f"  Failed: {result['failed']}")
+
+        # Get final stats
+        print("\nFinal embedding stats:")
+        stats = await get_embedding_stats()
+        print(f"  Total approved: {stats['total_approved']}")
+        print(f"  With embeddings: {stats['with_embeddings']}")
+        print(f"  Without embeddings: {stats['without_embeddings']}")
+        print(f"  Coverage: {stats['coverage_percent']}%")
+
+        return 0 if result["failed"] == 0 else 1
+
+    finally:
+        await client.disconnect()
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="Backfill embeddings for store agents")
+    parser.add_argument(
+        "--batch-size",
+        type=int,
+        default=100,
+        help="Number of embeddings to generate (default: 100)",
+    )
+    args = parser.parse_args()
+
+    sys.exit(asyncio.run(main(batch_size=args.batch_size)))

autogpt_platform/backend/backend/api/features/store/db.py

@@ -1,6 +1,5 @@
 import asyncio
 import logging
-import typing
 from datetime import datetime, timezone
 from typing import Literal
 
@@ -10,7 +9,7 @@ import prisma.errors
 import prisma.models
 import prisma.types
 
-from backend.data.db import query_raw_with_schema, transaction
+from backend.data.db import transaction
 from backend.data.graph import (
     GraphMeta,
     GraphModel,
@@ -30,6 +29,8 @@ from backend.util.settings import Settings
 
 from . import exceptions as store_exceptions
 from . import model as store_model
+from .embeddings import ensure_embedding
+from .hybrid_search import hybrid_search
 
 logger = logging.getLogger(__name__)
 settings = Settings()
@@ -56,122 +57,62 @@ async def get_store_agents(
         f"Getting store agents. featured={featured}, creators={creators}, sorted_by={sorted_by}, search={search_query}, category={category}, page={page}"
     )
 
+    search_used_hybrid = False
+    store_agents: list[store_model.StoreAgent] = []
+    total = 0
+    total_pages = 0
+
     try:
-        # If search_query is provided, use full-text search
+        # If search_query is provided, try hybrid search (embeddings + tsvector)
         if search_query:
-            offset = (page - 1) * page_size
+            try:
+                # Use hybrid search combining semantic and lexical signals
+                agents, total = await hybrid_search(
+                    query=search_query,
+                    featured=featured,
+                    creators=creators,
+                    category=category,
+                    sorted_by="relevance",  # Use hybrid scoring for relevance
+                    page=page,
+                    page_size=page_size,
+                )
+                search_used_hybrid = True
 
-            # Whitelist allowed order_by columns
-            ALLOWED_ORDER_BY = {
-                "rating": "rating DESC, rank DESC",
-                "runs": "runs DESC, rank DESC",
-                "name": "agent_name ASC, rank ASC",
-                "updated_at": "updated_at DESC, rank DESC",
-            }
+                # Convert hybrid search results (dict format)
+                total_pages = (total + page_size - 1) // page_size
+                store_agents: list[store_model.StoreAgent] = []
+                for agent in agents:
+                    try:
+                        store_agent = store_model.StoreAgent(
+                            slug=agent["slug"],
+                            agent_name=agent["agent_name"],
+                            agent_image=(
+                                agent["agent_image"][0] if agent["agent_image"] else ""
+                            ),
+                            creator=agent["creator_username"] or "Needs Profile",
+                            creator_avatar=agent["creator_avatar"] or "",
+                            sub_heading=agent["sub_heading"],
+                            description=agent["description"],
+                            runs=agent["runs"],
+                            rating=agent["rating"],
+                        )
+                        store_agents.append(store_agent)
+                    except Exception as e:
+                        logger.error(
+                            f"Error parsing Store agent from hybrid search results: {e}"
+                        )
+                        continue
 
-            # Validate and get order clause
-            if sorted_by and sorted_by in ALLOWED_ORDER_BY:
-                order_by_clause = ALLOWED_ORDER_BY[sorted_by]
-            else:
-                order_by_clause = "updated_at DESC, rank DESC"
+            except Exception as hybrid_error:
+                # If hybrid search fails (e.g., missing embeddings table),
+                # fallback to basic search logic below
+                logger.warning(
+                    f"Hybrid search failed, falling back to basic search: {hybrid_error}"
+                )
+                search_used_hybrid = False
 
-            # Build WHERE conditions and parameters list
-            where_parts: list[str] = []
-            params: list[typing.Any] = [search_query]  # $1 - search term
-            param_index = 2  # Start at $2 for next parameter
-
-            # Always filter for available agents
-            where_parts.append("is_available = true")
-
-            if featured:
-                where_parts.append("featured = true")
-
-            if creators and creators:
-                # Use ANY with array parameter
-                where_parts.append(f"creator_username = ANY(${param_index})")
-                params.append(creators)
-                param_index += 1
-
-            if category and category:
-                where_parts.append(f"${param_index} = ANY(categories)")
-                params.append(category)
-                param_index += 1
-
-            sql_where_clause: str = " AND ".join(where_parts) if where_parts else "1=1"
-
-            # Add pagination params
-            params.extend([page_size, offset])
-            limit_param = f"${param_index}"
-            offset_param = f"${param_index + 1}"
-
-            # Execute full-text search query with parameterized values
-            sql_query = f"""
-                SELECT
-                    slug,
-                    agent_name,
-                    agent_image,
-                    creator_username,
-                    creator_avatar,
-                    sub_heading,
-                    description,
-                    runs,
-                    rating,
-                    categories,
-                    featured,
-                    is_available,
-                    updated_at,
-                    ts_rank_cd(search, query) AS rank
-                FROM {{schema_prefix}}"StoreAgent",
-                    plainto_tsquery('english', $1) AS query
-                WHERE {sql_where_clause}
-                    AND search @@ query
-                ORDER BY {order_by_clause}
-                LIMIT {limit_param} OFFSET {offset_param}
-            """
-
-            # Count query for pagination - only uses search term parameter
-            count_query = f"""
-                SELECT COUNT(*) as count
-                FROM {{schema_prefix}}"StoreAgent",
-                    plainto_tsquery('english', $1) AS query
-                WHERE {sql_where_clause}
-                    AND search @@ query
-            """
-
-            # Execute both queries with parameters
-            agents = await query_raw_with_schema(sql_query, *params)
-
-            # For count, use params without pagination (last 2 params)
-            count_params = params[:-2]
-            count_result = await query_raw_with_schema(count_query, *count_params)
-
-            total = count_result[0]["count"] if count_result else 0
-            total_pages = (total + page_size - 1) // page_size
-
-            # Convert raw results to StoreAgent models
-            store_agents: list[store_model.StoreAgent] = []
-            for agent in agents:
-                try:
-                    store_agent = store_model.StoreAgent(
-                        slug=agent["slug"],
-                        agent_name=agent["agent_name"],
-                        agent_image=(
-                            agent["agent_image"][0] if agent["agent_image"] else ""
-                        ),
-                        creator=agent["creator_username"] or "Needs Profile",
-                        creator_avatar=agent["creator_avatar"] or "",
-                        sub_heading=agent["sub_heading"],
-                        description=agent["description"],
-                        runs=agent["runs"],
-                        rating=agent["rating"],
-                    )
-                    store_agents.append(store_agent)
-                except Exception as e:
-                    logger.error(f"Error parsing Store agent from search results: {e}")
-                    continue
-
-        else:
-            # Non-search query path (original logic)
+        if not search_used_hybrid:
+            # Fallback path - use basic search or no search
             where_clause: prisma.types.StoreAgentWhereInput = {"is_available": True}
             if featured:
                 where_clause["featured"] = featured
@@ -180,6 +121,14 @@ async def get_store_agents(
             if category:
                 where_clause["categories"] = {"has": category}
 
+            # Add basic text search if search_query provided but hybrid failed
+            if search_query:
+                where_clause["OR"] = [
+                    {"agent_name": {"contains": search_query, "mode": "insensitive"}},
+                    {"sub_heading": {"contains": search_query, "mode": "insensitive"}},
+                    {"description": {"contains": search_query, "mode": "insensitive"}},
+                ]
+
             order_by = []
             if sorted_by == "rating":
                 order_by.append({"rating": "desc"})
@@ -1600,6 +1549,22 @@ async def review_store_submission(
                     },
                 )
 
+            # Generate embedding for approved listing (non-blocking)
+            try:
+                await ensure_embedding(
+                    version_id=store_listing_version_id,
+                    name=store_listing_version.name,
+                    description=store_listing_version.description,
+                    sub_heading=store_listing_version.subHeading,
+                    categories=store_listing_version.categories or [],
+                )
+            except Exception as e:
+                # Don't fail approval if embedding generation fails
+                logger.warning(
+                    f"Failed to generate embedding for approved listing "
+                    f"{store_listing_version_id}: {e}"
+                )
+
         # If rejecting an approved agent, update the StoreListing accordingly
         if is_rejecting_approved:
             # Check if there are other approved versions

autogpt_platform/backend/backend/api/features/store/embeddings.py

@@ -0,0 +1,533 @@
+"""
+Unified Content Embeddings Service
+
+Handles generation and storage of OpenAI embeddings for all content types
+(store listings, blocks, documentation, library agents) to enable semantic/hybrid search.
+"""
+
+import asyncio
+import logging
+from typing import Any
+
+import prisma
+from openai import OpenAI
+from prisma.enums import ContentType
+
+from backend.util.json import dumps
+from backend.util.settings import Settings
+
+logger = logging.getLogger(__name__)
+
+# OpenAI embedding model configuration
+EMBEDDING_MODEL = "text-embedding-3-small"
+EMBEDDING_DIM = 1536
+
+
+def build_searchable_text(
+    name: str,
+    description: str,
+    sub_heading: str,
+    categories: list[str],
+) -> str:
+    """
+    Build searchable text from listing version fields.
+
+    Combines relevant fields into a single string for embedding.
+    """
+    parts = []
+
+    # Name is important - include it
+    if name:
+        parts.append(name)
+
+    # Sub-heading provides context
+    if sub_heading:
+        parts.append(sub_heading)
+
+    # Description is the main content
+    if description:
+        parts.append(description)
+
+    # Categories help with semantic matching
+    if categories:
+        parts.append(" ".join(categories))
+
+    return " ".join(parts)
+
+
+async def generate_embedding(text: str) -> list[float] | None:
+    """
+    Generate embedding for text using OpenAI API.
+
+    Returns None if embedding generation fails.
+    """
+    try:
+        settings = Settings()
+        api_key = settings.secrets.openai_internal_api_key
+        if not api_key:
+            logger.warning("openai_internal_api_key not set, cannot generate embedding")
+            return None
+
+        client = OpenAI(api_key=api_key)
+
+        # Truncate text to avoid token limits (~32k chars for safety)
+        truncated_text = text[:32000]
+
+        response = client.embeddings.create(
+            model=EMBEDDING_MODEL,
+            input=truncated_text,
+        )
+
+        embedding = response.data[0].embedding
+        logger.debug(f"Generated embedding with {len(embedding)} dimensions")
+        return embedding
+
+    except Exception as e:
+        logger.error(f"Failed to generate embedding: {e}")
+        return None
+
+
+async def store_embedding(
+    version_id: str,
+    embedding: list[float],
+    tx: prisma.Prisma | None = None,
+) -> bool:
+    """
+    Store embedding in the database.
+
+    BACKWARD COMPATIBILITY: Maintained for existing store listing usage.
+    Uses raw SQL since Prisma doesn't natively support pgvector.
+    """
+    return await store_content_embedding(
+        content_type=ContentType.STORE_AGENT,
+        content_id=version_id,
+        embedding=embedding,
+        searchable_text="",  # Will be populated from existing data
+        metadata=None,
+        user_id=None,  # Store agents are public
+        tx=tx,
+    )
+
+
+async def store_content_embedding(
+    content_type: ContentType,
+    content_id: str,
+    embedding: list[float],
+    searchable_text: str,
+    metadata: dict | None = None,
+    user_id: str | None = None,
+    tx: prisma.Prisma | None = None,
+) -> bool:
+    """
+    Store embedding in the unified content embeddings table.
+
+    New function for unified content embedding storage.
+    Uses raw SQL since Prisma doesn't natively support pgvector.
+    """
+    try:
+        client = tx if tx else prisma.get_client()
+
+        # Convert embedding to PostgreSQL vector format
+        embedding_str = "[" + ",".join(str(x) for x in embedding) + "]"
+        metadata_json = dumps(metadata or {})
+
+        # Upsert the embedding
+        await client.execute_raw(
+            """
+            INSERT INTO platform."UnifiedContentEmbedding" (
+                "contentType", "contentId", "userId", "embedding", "searchableText", "metadata", "createdAt", "updatedAt"
+            )
+            VALUES ($1, $2, $3, $4::vector, $5, $6::jsonb, NOW(), NOW())
+            ON CONFLICT ("contentType", "contentId", "userId")
+            DO UPDATE SET
+                "embedding" = $4::vector,
+                "searchableText" = $5,
+                "metadata" = $6::jsonb,
+                "updatedAt" = NOW()
+            """,
+            content_type,
+            content_id,
+            user_id,
+            embedding_str,
+            searchable_text,
+            metadata_json,
+        )
+
+        logger.info(f"Stored embedding for {content_type}:{content_id}")
+        return True
+
+    except Exception as e:
+        logger.error(f"Failed to store embedding for {content_type}:{content_id}: {e}")
+        return False
+
+
+async def get_embedding(version_id: str) -> dict[str, Any] | None:
+    """
+    Retrieve embedding record for a listing version.
+
+    BACKWARD COMPATIBILITY: Maintained for existing store listing usage.
+    Returns dict with storeListingVersionId, embedding, timestamps or None if not found.
+    """
+    result = await get_content_embedding(
+        ContentType.STORE_AGENT, version_id, user_id=None
+    )
+    if result:
+        # Transform to old format for backward compatibility
+        return {
+            "storeListingVersionId": result["contentId"],
+            "embedding": result["embedding"],
+            "createdAt": result["createdAt"],
+            "updatedAt": result["updatedAt"],
+        }
+    return None
+
+
+async def get_content_embedding(
+    content_type: ContentType, content_id: str, user_id: str | None = None
+) -> dict[str, Any] | None:
+    """
+    Retrieve embedding record for any content type.
+
+    New function for unified content embedding retrieval.
+    Returns dict with contentType, contentId, embedding, timestamps or None if not found.
+    """
+    try:
+        client = prisma.get_client()
+
+        result = await client.query_raw(
+            """
+            SELECT
+                "contentType",
+                "contentId",
+                "userId",
+                "embedding"::text as "embedding",
+                "searchableText",
+                "metadata",
+                "createdAt",
+                "updatedAt"
+            FROM platform."UnifiedContentEmbedding"
+            WHERE "contentType" = $1 AND "contentId" = $2 AND ("userId" = $3 OR ($3 IS NULL AND "userId" IS NULL))
+            """,
+            content_type,
+            content_id,
+            user_id,
+        )
+
+        if result and len(result) > 0:
+            return result[0]
+        return None
+
+    except Exception as e:
+        logger.error(f"Failed to get embedding for {content_type}:{content_id}: {e}")
+        return None
+
+
+async def ensure_embedding(
+    version_id: str,
+    name: str,
+    description: str,
+    sub_heading: str,
+    categories: list[str],
+    force: bool = False,
+    tx: prisma.Prisma | None = None,
+) -> bool:
+    """
+    Ensure an embedding exists for the listing version.
+
+    Creates embedding if missing. Use force=True to regenerate.
+    Backward-compatible wrapper for store listings.
+
+    Args:
+        version_id: The StoreListingVersion ID
+        name: Agent name
+        description: Agent description
+        sub_heading: Agent sub-heading
+        categories: Agent categories
+        force: Force regeneration even if embedding exists
+        tx: Optional transaction client
+
+    Returns:
+        True if embedding exists/was created, False on failure
+    """
+    try:
+        # Check if embedding already exists
+        if not force:
+            existing = await get_embedding(version_id)
+            if existing and existing.get("embedding"):
+                logger.debug(f"Embedding for version {version_id} already exists")
+                return True
+
+        # Build searchable text for embedding
+        searchable_text = build_searchable_text(
+            name, description, sub_heading, categories
+        )
+
+        # Generate new embedding
+        embedding = await generate_embedding(searchable_text)
+        if embedding is None:
+            logger.warning(f"Could not generate embedding for version {version_id}")
+            return False
+
+        # Store the embedding with metadata using new function
+        metadata = {
+            "name": name,
+            "subHeading": sub_heading,
+            "categories": categories,
+        }
+        return await store_content_embedding(
+            content_type=ContentType.STORE_AGENT,
+            content_id=version_id,
+            embedding=embedding,
+            searchable_text=searchable_text,
+            metadata=metadata,
+            user_id=None,  # Store agents are public
+            tx=tx,
+        )
+
+    except Exception as e:
+        logger.error(f"Failed to ensure embedding for version {version_id}: {e}")
+        return False
+
+
+async def delete_embedding(version_id: str) -> bool:
+    """
+    Delete embedding for a listing version.
+
+    BACKWARD COMPATIBILITY: Maintained for existing store listing usage.
+    Note: This is usually handled automatically by CASCADE delete,
+    but provided for manual cleanup if needed.
+    """
+    return await delete_content_embedding(ContentType.STORE_AGENT, version_id)
+
+
+async def delete_content_embedding(content_type: ContentType, content_id: str) -> bool:
+    """
+    Delete embedding for any content type.
+
+    New function for unified content embedding deletion.
+    Note: This is usually handled automatically by CASCADE delete,
+    but provided for manual cleanup if needed.
+    """
+    try:
+        client = prisma.get_client()
+
+        await client.execute_raw(
+            """
+            DELETE FROM platform."UnifiedContentEmbedding"
+            WHERE "contentType" = $1 AND "contentId" = $2
+            """,
+            content_type,
+            content_id,
+        )
+
+        logger.info(f"Deleted embedding for {content_type}:{content_id}")
+        return True
+
+    except Exception as e:
+        logger.error(f"Failed to delete embedding for {content_type}:{content_id}: {e}")
+        return False
+
+
+async def get_embedding_stats() -> dict[str, Any]:
+    """
+    Get statistics about embedding coverage.
+
+    Returns counts of:
+    - Total approved listing versions
+    - Versions with embeddings
+    - Versions without embeddings
+    """
+    try:
+        client = prisma.get_client()
+
+        # Count approved versions
+        approved_result = await client.query_raw(
+            """
+            SELECT COUNT(*) as count
+            FROM platform."StoreListingVersion"
+            WHERE "submissionStatus" = 'APPROVED'
+            AND "isDeleted" = false
+            """
+        )
+        total_approved = approved_result[0]["count"] if approved_result else 0
+
+        # Count versions with embeddings
+        embedded_result = await client.query_raw(
+            """
+            SELECT COUNT(*) as count
+            FROM platform."StoreListingVersion" slv
+            JOIN platform."UnifiedContentEmbedding" uce ON slv.id = uce."contentId" AND uce."contentType" = 'STORE_AGENT'
+            WHERE slv."submissionStatus" = 'APPROVED'
+            AND slv."isDeleted" = false
+            """
+        )
+        with_embeddings = embedded_result[0]["count"] if embedded_result else 0
+
+        return {
+            "total_approved": total_approved,
+            "with_embeddings": with_embeddings,
+            "without_embeddings": total_approved - with_embeddings,
+            "coverage_percent": (
+                round(with_embeddings / total_approved * 100, 1)
+                if total_approved > 0
+                else 0
+            ),
+        }
+
+    except Exception as e:
+        logger.error(f"Failed to get embedding stats: {e}")
+        return {
+            "total_approved": 0,
+            "with_embeddings": 0,
+            "without_embeddings": 0,
+            "coverage_percent": 0,
+            "error": str(e),
+        }
+
+
+async def backfill_missing_embeddings(batch_size: int = 10) -> dict[str, Any]:
+    """
+    Generate embeddings for approved listings that don't have them.
+
+    Args:
+        batch_size: Number of embeddings to generate in one call
+
+    Returns:
+        Dict with success/failure counts
+    """
+    try:
+        client = prisma.get_client()
+
+        # Find approved versions without embeddings
+        missing = await client.query_raw(
+            """
+            SELECT
+                slv.id,
+                slv.name,
+                slv.description,
+                slv."subHeading",
+                slv.categories
+            FROM platform."StoreListingVersion" slv
+            LEFT JOIN platform."UnifiedContentEmbedding" uce
+                ON slv.id = uce."contentId" AND uce."contentType" = 'STORE_AGENT'
+            WHERE slv."submissionStatus" = 'APPROVED'
+            AND slv."isDeleted" = false
+            AND uce."contentId" IS NULL
+            LIMIT $1
+            """,
+            batch_size,
+        )
+
+        if not missing:
+            return {
+                "processed": 0,
+                "success": 0,
+                "failed": 0,
+                "message": "No missing embeddings",
+            }
+
+        # Process embeddings concurrently for better performance
+        embedding_tasks = [
+            ensure_embedding(
+                version_id=row["id"],
+                name=row["name"],
+                description=row["description"],
+                sub_heading=row["subHeading"],
+                categories=row["categories"] or [],
+            )
+            for row in missing
+        ]
+
+        results = await asyncio.gather(*embedding_tasks, return_exceptions=True)
+
+        success = sum(1 for result in results if result is True)
+        failed = len(results) - success
+
+        return {
+            "processed": len(missing),
+            "success": success,
+            "failed": failed,
+            "message": f"Backfilled {success} embeddings, {failed} failed",
+        }
+
+    except Exception as e:
+        logger.error(f"Failed to backfill embeddings: {e}")
+        return {
+            "processed": 0,
+            "success": 0,
+            "failed": 0,
+            "error": str(e),
+        }
+
+
+async def embed_query(query: str) -> list[float] | None:
+    """
+    Generate embedding for a search query.
+
+    Same as generate_embedding but with clearer intent.
+    """
+    return await generate_embedding(query)
+
+
+def embedding_to_vector_string(embedding: list[float]) -> str:
+    """Convert embedding list to PostgreSQL vector string format."""
+    return "[" + ",".join(str(x) for x in embedding) + "]"
+
+
+async def ensure_content_embedding(
+    content_type: ContentType,
+    content_id: str,
+    searchable_text: str,
+    metadata: dict | None = None,
+    user_id: str | None = None,
+    force: bool = False,
+    tx: prisma.Prisma | None = None,
+) -> bool:
+    """
+    Ensure an embedding exists for any content type.
+
+    Generic function for creating embeddings for store agents, blocks, docs, etc.
+
+    Args:
+        content_type: ContentType enum value (STORE_AGENT, BLOCK, etc.)
+        content_id: Unique identifier for the content
+        searchable_text: Combined text for embedding generation
+        metadata: Optional metadata to store with embedding
+        force: Force regeneration even if embedding exists
+        tx: Optional transaction client
+
+    Returns:
+        True if embedding exists/was created, False on failure
+    """
+    try:
+        # Check if embedding already exists
+        if not force:
+            existing = await get_content_embedding(content_type, content_id, user_id)
+            if existing and existing.get("embedding"):
+                logger.debug(
+                    f"Embedding for {content_type}:{content_id} already exists"
+                )
+                return True
+
+        # Generate new embedding
+        embedding = await generate_embedding(searchable_text)
+        if embedding is None:
+            logger.warning(
+                f"Could not generate embedding for {content_type}:{content_id}"
+            )
+            return False
+
+        # Store the embedding
+        return await store_content_embedding(
+            content_type=content_type,
+            content_id=content_id,
+            embedding=embedding,
+            searchable_text=searchable_text,
+            metadata=metadata or {},
+            user_id=user_id,
+            tx=tx,
+        )
+
+    except Exception as e:
+        logger.error(f"Failed to ensure embedding for {content_type}:{content_id}: {e}")
+        return False

autogpt_platform/backend/backend/api/features/store/embeddings_test.py

@@ -0,0 +1,359 @@
+from unittest.mock import MagicMock, patch
+
+import prisma
+import pytest
+from prisma import Prisma
+from prisma.enums import ContentType
+
+from backend.api.features.store import embeddings
+
+
+@pytest.fixture(autouse=True)
+async def setup_prisma():
+    """Setup Prisma client for tests."""
+    try:
+        Prisma()
+    except prisma.errors.ClientAlreadyRegisteredError:
+        pass
+    yield
+
+
+@pytest.mark.asyncio(loop_scope="session")
+async def test_build_searchable_text():
+    """Test searchable text building from listing fields."""
+    result = embeddings.build_searchable_text(
+        name="AI Assistant",
+        description="A helpful AI assistant for productivity",
+        sub_heading="Boost your productivity",
+        categories=["AI", "Productivity"],
+    )
+
+    expected = "AI Assistant Boost your productivity A helpful AI assistant for productivity AI Productivity"
+    assert result == expected
+
+
+@pytest.mark.asyncio(loop_scope="session")
+async def test_build_searchable_text_empty_fields():
+    """Test searchable text building with empty fields."""
+    result = embeddings.build_searchable_text(
+        name="", description="Test description", sub_heading="", categories=[]
+    )
+
+    assert result == "Test description"
+
+
+@pytest.mark.asyncio(loop_scope="session")
+@patch("backend.api.features.store.embeddings.OpenAI")
+async def test_generate_embedding_success(mock_openai_class):
+    """Test successful embedding generation."""
+    # Mock OpenAI response
+    mock_client = MagicMock()
+    mock_response = MagicMock()
+    mock_response.data = [MagicMock()]
+    mock_response.data[0].embedding = [0.1, 0.2, 0.3] * 512  # 1536 dimensions
+    mock_client.embeddings.create.return_value = mock_response
+    mock_openai_class.return_value = mock_client
+
+    with patch("backend.api.features.store.embeddings.Settings") as mock_settings:
+        mock_settings.return_value.secrets.openai_internal_api_key = "test-key"
+
+        result = await embeddings.generate_embedding("test text")
+
+        assert result is not None
+        assert len(result) == 1536
+        assert result[0] == 0.1
+
+        mock_client.embeddings.create.assert_called_once_with(
+            model="text-embedding-3-small", input="test text"
+        )
+
+
+@pytest.mark.asyncio(loop_scope="session")
+@patch("backend.api.features.store.embeddings.OpenAI")
+async def test_generate_embedding_no_api_key(mock_openai_class):
+    """Test embedding generation without API key."""
+    with patch("backend.api.features.store.embeddings.Settings") as mock_settings:
+        mock_settings.return_value.secrets.openai_internal_api_key = ""
+
+        result = await embeddings.generate_embedding("test text")
+
+        assert result is None
+        mock_openai_class.assert_not_called()
+
+
+@pytest.mark.asyncio(loop_scope="session")
+@patch("backend.api.features.store.embeddings.OpenAI")
+async def test_generate_embedding_api_error(mock_openai_class):
+    """Test embedding generation with API error."""
+    mock_client = MagicMock()
+    mock_client.embeddings.create.side_effect = Exception("API Error")
+    mock_openai_class.return_value = mock_client
+
+    with patch("backend.api.features.store.embeddings.Settings") as mock_settings:
+        mock_settings.return_value.secrets.openai_internal_api_key = "test-key"
+
+        result = await embeddings.generate_embedding("test text")
+
+        assert result is None
+
+
+@pytest.mark.asyncio(loop_scope="session")
+@patch("backend.api.features.store.embeddings.OpenAI")
+async def test_generate_embedding_text_truncation(mock_openai_class):
+    """Test that long text is properly truncated."""
+    mock_client = MagicMock()
+    mock_response = MagicMock()
+    mock_response.data = [MagicMock()]
+    mock_response.data[0].embedding = [0.1] * 1536
+    mock_client.embeddings.create.return_value = mock_response
+    mock_openai_class.return_value = mock_client
+
+    # Create text longer than 32k chars
+    long_text = "a" * 35000
+
+    with patch("backend.api.features.store.embeddings.Settings") as mock_settings:
+        mock_settings.return_value.secrets.openai_internal_api_key = "test-key"
+
+        await embeddings.generate_embedding(long_text)
+
+        # Verify truncated text was sent to API
+        call_args = mock_client.embeddings.create.call_args
+        assert len(call_args.kwargs["input"]) == 32000
+
+
+@pytest.mark.asyncio(loop_scope="session")
+async def test_store_embedding_success(mocker):
+    """Test successful embedding storage."""
+    mock_client = mocker.AsyncMock()
+    mock_client.execute_raw = mocker.AsyncMock()
+
+    embedding = [0.1, 0.2, 0.3]
+
+    result = await embeddings.store_embedding(
+        version_id="test-version-id", embedding=embedding, tx=mock_client
+    )
+
+    assert result is True
+    mock_client.execute_raw.assert_called_once()
+    call_args = mock_client.execute_raw.call_args[0]
+    assert "test-version-id" in call_args
+    assert "[0.1,0.2,0.3]" in call_args
+    assert None in call_args  # userId should be None for store agents
+
+
+@pytest.mark.asyncio(loop_scope="session")
+async def test_store_embedding_database_error(mocker):
+    """Test embedding storage with database error."""
+    mock_client = mocker.AsyncMock()
+    mock_client.execute_raw.side_effect = Exception("Database error")
+
+    embedding = [0.1, 0.2, 0.3]
+
+    result = await embeddings.store_embedding(
+        version_id="test-version-id", embedding=embedding, tx=mock_client
+    )
+
+    assert result is False
+
+
+@pytest.mark.asyncio(loop_scope="session")
+async def test_get_embedding_success(mocker):
+    """Test successful embedding retrieval."""
+    mock_client = mocker.AsyncMock()
+    mock_result = [
+        {
+            "contentType": "STORE_AGENT",
+            "contentId": "test-version-id",
+            "embedding": "[0.1,0.2,0.3]",
+            "searchableText": "Test text",
+            "metadata": {},
+            "createdAt": "2024-01-01T00:00:00Z",
+            "updatedAt": "2024-01-01T00:00:00Z",
+        }
+    ]
+    mock_client.query_raw.return_value = mock_result
+
+    with patch("prisma.get_client", return_value=mock_client):
+        result = await embeddings.get_embedding("test-version-id")
+
+        assert result is not None
+        assert result["storeListingVersionId"] == "test-version-id"
+        assert result["embedding"] == "[0.1,0.2,0.3]"
+
+
+@pytest.mark.asyncio(loop_scope="session")
+async def test_get_embedding_not_found(mocker):
+    """Test embedding retrieval when not found."""
+    mock_client = mocker.AsyncMock()
+    mock_client.query_raw.return_value = []
+
+    with patch("prisma.get_client", return_value=mock_client):
+        result = await embeddings.get_embedding("test-version-id")
+
+        assert result is None
+
+
+@pytest.mark.asyncio(loop_scope="session")
+@patch("backend.api.features.store.embeddings.generate_embedding")
+@patch("backend.api.features.store.embeddings.store_embedding")
+@patch("backend.api.features.store.embeddings.get_embedding")
+async def test_ensure_embedding_already_exists(mock_get, mock_store, mock_generate):
+    """Test ensure_embedding when embedding already exists."""
+    mock_get.return_value = {"embedding": "[0.1,0.2,0.3]"}
+
+    result = await embeddings.ensure_embedding(
+        version_id="test-id",
+        name="Test",
+        description="Test description",
+        sub_heading="Test heading",
+        categories=["test"],
+    )
+
+    assert result is True
+    mock_generate.assert_not_called()
+    mock_store.assert_not_called()
+
+
+@pytest.mark.asyncio(loop_scope="session")
+@patch("backend.api.features.store.embeddings.generate_embedding")
+@patch("backend.api.features.store.embeddings.store_content_embedding")
+@patch("backend.api.features.store.embeddings.get_embedding")
+async def test_ensure_embedding_create_new(mock_get, mock_store, mock_generate):
+    """Test ensure_embedding creating new embedding."""
+    mock_get.return_value = None
+    mock_generate.return_value = [0.1, 0.2, 0.3]
+    mock_store.return_value = True
+
+    result = await embeddings.ensure_embedding(
+        version_id="test-id",
+        name="Test",
+        description="Test description",
+        sub_heading="Test heading",
+        categories=["test"],
+    )
+
+    assert result is True
+    mock_generate.assert_called_once_with("Test Test heading Test description test")
+    mock_store.assert_called_once_with(
+        content_type=ContentType.STORE_AGENT,
+        content_id="test-id",
+        embedding=[0.1, 0.2, 0.3],
+        searchable_text="Test Test heading Test description test",
+        metadata={"name": "Test", "subHeading": "Test heading", "categories": ["test"]},
+        user_id=None,
+        tx=None,
+    )
+
+
+@pytest.mark.asyncio(loop_scope="session")
+@patch("backend.api.features.store.embeddings.generate_embedding")
+@patch("backend.api.features.store.embeddings.get_embedding")
+async def test_ensure_embedding_generation_fails(mock_get, mock_generate):
+    """Test ensure_embedding when generation fails."""
+    mock_get.return_value = None
+    mock_generate.return_value = None
+
+    result = await embeddings.ensure_embedding(
+        version_id="test-id",
+        name="Test",
+        description="Test description",
+        sub_heading="Test heading",
+        categories=["test"],
+    )
+
+    assert result is False
+
+
+@pytest.mark.asyncio(loop_scope="session")
+async def test_get_embedding_stats(mocker):
+    """Test embedding statistics retrieval."""
+    mock_client = mocker.AsyncMock()
+
+    # Mock approved count query
+    mock_approved_result = [{"count": 100}]
+    # Mock embedded count query
+    mock_embedded_result = [{"count": 75}]
+
+    mock_client.query_raw.side_effect = [mock_approved_result, mock_embedded_result]
+
+    with patch("prisma.get_client", return_value=mock_client):
+        result = await embeddings.get_embedding_stats()
+
+        assert result["total_approved"] == 100
+        assert result["with_embeddings"] == 75
+        assert result["without_embeddings"] == 25
+        assert result["coverage_percent"] == 75.0
+
+
+@pytest.mark.asyncio(loop_scope="session")
+@patch("backend.api.features.store.embeddings.ensure_embedding")
+async def test_backfill_missing_embeddings_success(mock_ensure, mocker):
+    """Test backfill with successful embedding generation."""
+    mock_client = mocker.AsyncMock()
+
+    # Mock missing embeddings query
+    mock_missing = [
+        {
+            "id": "version-1",
+            "name": "Agent 1",
+            "description": "Description 1",
+            "subHeading": "Heading 1",
+            "categories": ["AI"],
+        },
+        {
+            "id": "version-2",
+            "name": "Agent 2",
+            "description": "Description 2",
+            "subHeading": "Heading 2",
+            "categories": ["Productivity"],
+        },
+    ]
+    mock_client.query_raw.return_value = mock_missing
+
+    # Mock ensure_embedding to succeed for first, fail for second
+    mock_ensure.side_effect = [True, False]
+
+    with patch("prisma.get_client", return_value=mock_client):
+        result = await embeddings.backfill_missing_embeddings(batch_size=5)
+
+        assert result["processed"] == 2
+        assert result["success"] == 1
+        assert result["failed"] == 1
+        assert mock_ensure.call_count == 2
+
+
+@pytest.mark.asyncio(loop_scope="session")
+async def test_backfill_missing_embeddings_no_missing(mocker):
+    """Test backfill when no embeddings are missing."""
+    mock_client = mocker.AsyncMock()
+    mock_client.query_raw.return_value = []
+
+    with patch("prisma.get_client", return_value=mock_client):
+        result = await embeddings.backfill_missing_embeddings(batch_size=5)
+
+        assert result["processed"] == 0
+        assert result["success"] == 0
+        assert result["failed"] == 0
+        assert result["message"] == "No missing embeddings"
+
+
+@pytest.mark.asyncio(loop_scope="session")
+async def test_embedding_to_vector_string():
+    """Test embedding to PostgreSQL vector string conversion."""
+    embedding = [0.1, 0.2, 0.3, -0.4]
+    result = embeddings.embedding_to_vector_string(embedding)
+    assert result == "[0.1,0.2,0.3,-0.4]"
+
+
+@pytest.mark.asyncio(loop_scope="session")
+async def test_embed_query():
+    """Test embed_query function (alias for generate_embedding)."""
+    with patch(
+        "backend.api.features.store.embeddings.generate_embedding"
+    ) as mock_generate:
+        mock_generate.return_value = [0.1, 0.2, 0.3]
+
+        result = await embeddings.embed_query("test query")
+
+        assert result == [0.1, 0.2, 0.3]
+        mock_generate.assert_called_once_with("test query")

autogpt_platform/backend/backend/api/features/store/hybrid_search.py

@@ -0,0 +1,377 @@
+"""
+Hybrid Search for Store Agents
+
+Combines semantic (embedding) search with lexical (tsvector) search
+for improved relevance in marketplace agent discovery.
+"""
+
+import logging
+from dataclasses import dataclass
+from datetime import datetime
+from typing import Any, Literal
+
+from backend.api.features.store.embeddings import (
+    embed_query,
+    embedding_to_vector_string,
+)
+from backend.data.db import query_raw_with_schema
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class HybridSearchWeights:
+    """Weights for combining search signals."""
+
+    semantic: float = 0.35  # Embedding cosine similarity
+    lexical: float = 0.35  # tsvector ts_rank_cd score
+    category: float = 0.20  # Category match boost
+    recency: float = 0.10  # Newer agents ranked higher
+
+
+DEFAULT_WEIGHTS = HybridSearchWeights()
+
+# Minimum relevance score threshold - agents below this are filtered out
+# With weights (0.35 semantic + 0.35 lexical + 0.20 category + 0.10 recency):
+# - 0.20 means at least ~50% semantic match OR strong lexical match required
+# - Ensures only genuinely relevant results are returned
+# - Recency alone (0.10 max) won't pass the threshold
+DEFAULT_MIN_SCORE = 0.20
+
+
+@dataclass
+class HybridSearchResult:
+    """A single search result with score breakdown."""
+
+    slug: str
+    agent_name: str
+    agent_image: str
+    creator_username: str
+    creator_avatar: str
+    sub_heading: str
+    description: str
+    runs: int
+    rating: float
+    categories: list[str]
+    featured: bool
+    is_available: bool
+    updated_at: datetime
+
+    # Score breakdown (for debugging/tuning)
+    combined_score: float
+    semantic_score: float = 0.0
+    lexical_score: float = 0.0
+    category_score: float = 0.0
+    recency_score: float = 0.0
+
+
+async def hybrid_search(
+    query: str,
+    featured: bool = False,
+    creators: list[str] | None = None,
+    category: str | None = None,
+    sorted_by: (
+        Literal["relevance", "rating", "runs", "name", "updated_at"] | None
+    ) = None,
+    page: int = 1,
+    page_size: int = 20,
+    weights: HybridSearchWeights | None = None,
+    min_score: float | None = None,
+) -> tuple[list[dict[str, Any]], int]:
+    """
+    Perform hybrid search combining semantic and lexical signals.
+
+    Args:
+        query: Search query string
+        featured: Filter for featured agents only
+        creators: Filter by creator usernames
+        category: Filter by category
+        sorted_by: Sort order (relevance uses hybrid scoring)
+        page: Page number (1-indexed)
+        page_size: Results per page
+        weights: Custom weights for search signals
+        min_score: Minimum relevance score threshold (0-1). Results below
+                   this score are filtered out. Defaults to DEFAULT_MIN_SCORE.
+
+    Returns:
+        Tuple of (results list, total count). Returns empty list if no
+        results meet the minimum relevance threshold.
+    """
+    if weights is None:
+        weights = DEFAULT_WEIGHTS
+    if min_score is None:
+        min_score = DEFAULT_MIN_SCORE
+
+    offset = (page - 1) * page_size
+
+    # Generate query embedding
+    query_embedding = await embed_query(query)
+
+    # Build WHERE clause conditions
+    where_parts: list[str] = ["sa.is_available = true"]
+    params: list[Any] = []
+    param_index = 1
+
+    # Add search query for lexical matching
+    params.append(query)
+    query_param = f"${param_index}"
+    param_index += 1
+
+    # Add lowercased query for category matching
+    params.append(query.lower())
+    query_lower_param = f"${param_index}"
+    param_index += 1
+
+    if featured:
+        where_parts.append("sa.featured = true")
+
+    if creators:
+        where_parts.append(f"sa.creator_username = ANY(${param_index})")
+        params.append(creators)
+        param_index += 1
+
+    if category:
+        where_parts.append(f"${param_index} = ANY(sa.categories)")
+        params.append(category)
+        param_index += 1
+
+    where_clause = " AND ".join(where_parts)
+
+    # Determine if we can use hybrid search (have query embedding)
+    use_hybrid = query_embedding is not None
+
+    if use_hybrid:
+        # Add embedding parameter
+        embedding_str = embedding_to_vector_string(query_embedding)
+        params.append(embedding_str)
+        embedding_param = f"${param_index}"
+        param_index += 1
+
+        # Optimized hybrid search query:
+        # 1. Direct join to UnifiedContentEmbedding via contentId=storeListingVersionId (no redundant JOINs)
+        # 2. UNION ALL approach to enable index usage for both lexical and semantic branches
+        # 3. COUNT(*) OVER() to get total count in single query
+        # 4. Simplified category matching with array_to_string
+        sql_query = f"""
+            WITH candidates AS (
+                -- Lexical matches (uses GIN index on search column)
+                SELECT DISTINCT sa."storeListingVersionId"
+                FROM {{schema_prefix}}"StoreAgent" sa
+                WHERE {where_clause}
+                AND sa.search @@ plainto_tsquery('english', {query_param})
+
+                UNION
+
+                -- Semantic matches (uses HNSW index on embedding)
+                SELECT DISTINCT sa."storeListingVersionId"
+                FROM {{schema_prefix}}"StoreAgent" sa
+                INNER JOIN {{schema_prefix}}"UnifiedContentEmbedding" uce
+                    ON sa."storeListingVersionId" = uce."contentId" AND uce."contentType" = 'STORE_AGENT'
+                WHERE {where_clause}
+            ),
+            search_scores AS (
+                SELECT
+                    sa.slug,
+                    sa.agent_name,
+                    sa.agent_image,
+                    sa.creator_username,
+                    sa.creator_avatar,
+                    sa.sub_heading,
+                    sa.description,
+                    sa.runs,
+                    sa.rating,
+                    sa.categories,
+                    sa.featured,
+                    sa.is_available,
+                    sa.updated_at,
+                    -- Semantic score: cosine similarity (1 - distance)
+                    COALESCE(1 - (uce.embedding <=> {embedding_param}::vector), 0) as semantic_score,
+                    -- Lexical score: ts_rank_cd (will be normalized later)
+                    COALESCE(ts_rank_cd(sa.search, plainto_tsquery('english', {query_param})), 0) as lexical_raw,
+                    -- Category match: check if query appears in any category
+                    CASE
+                        WHEN LOWER(array_to_string(sa.categories, ' ')) LIKE '%' || {query_lower_param} || '%'
+                        THEN 1.0
+                        ELSE 0.0
+                    END as category_score,
+                    -- Recency score: exponential decay over 90 days
+                    EXP(-EXTRACT(EPOCH FROM (NOW() - sa.updated_at)) / (90 * 24 * 3600)) as recency_score
+                FROM candidates c
+                INNER JOIN {{schema_prefix}}"StoreAgent" sa
+                    ON c."storeListingVersionId" = sa."storeListingVersionId"
+                LEFT JOIN {{schema_prefix}}"UnifiedContentEmbedding" uce
+                    ON sa."storeListingVersionId" = uce."contentId" AND uce."contentType" = 'STORE_AGENT'
+            ),
+            normalized AS (
+                SELECT
+                    *,
+                    -- Normalize lexical score by max in result set
+                    CASE
+                        WHEN MAX(lexical_raw) OVER () > 0
+                        THEN lexical_raw / MAX(lexical_raw) OVER ()
+                        ELSE 0
+                    END as lexical_score
+                FROM search_scores
+            ),
+            scored AS (
+                SELECT
+                    slug,
+                    agent_name,
+                    agent_image,
+                    creator_username,
+                    creator_avatar,
+                    sub_heading,
+                    description,
+                    runs,
+                    rating,
+                    categories,
+                    featured,
+                    is_available,
+                    updated_at,
+                    semantic_score,
+                    lexical_score,
+                    category_score,
+                    recency_score,
+                    (
+                        {weights.semantic} * semantic_score +
+                        {weights.lexical} * lexical_score +
+                        {weights.category} * category_score +
+                        {weights.recency} * recency_score
+                    ) as combined_score
+                FROM normalized
+            ),
+            filtered AS (
+                SELECT
+                    *,
+                    COUNT(*) OVER () as total_count
+                FROM scored
+                WHERE combined_score >= {min_score}
+            )
+            SELECT * FROM filtered
+            ORDER BY combined_score DESC
+            LIMIT ${param_index} OFFSET ${param_index + 1}
+        """
+
+        # Add pagination params
+        params.extend([page_size, offset])
+
+    else:
+        # Fallback to lexical-only search (existing behavior)
+        logger.warning("Falling back to lexical-only search (no query embedding)")
+
+        sql_query = f"""
+            WITH lexical_scores AS (
+                SELECT
+                    slug,
+                    agent_name,
+                    agent_image,
+                    creator_username,
+                    creator_avatar,
+                    sub_heading,
+                    description,
+                    runs,
+                    rating,
+                    categories,
+                    featured,
+                    is_available,
+                    updated_at,
+                    0.0 as semantic_score,
+                    ts_rank_cd(search, plainto_tsquery('english', {query_param})) as lexical_raw,
+                    CASE
+                        WHEN LOWER(array_to_string(categories, ' ')) LIKE '%' || {query_lower_param} || '%'
+                        THEN 1.0
+                        ELSE 0.0
+                    END as category_score,
+                    EXP(-EXTRACT(EPOCH FROM (NOW() - updated_at)) / (90 * 24 * 3600)) as recency_score
+                FROM {{schema_prefix}}"StoreAgent" sa
+                WHERE {where_clause}
+                AND search @@ plainto_tsquery('english', {query_param})
+            ),
+            normalized AS (
+                SELECT
+                    *,
+                    CASE
+                        WHEN MAX(lexical_raw) OVER () > 0
+                        THEN lexical_raw / MAX(lexical_raw) OVER ()
+                        ELSE 0
+                    END as lexical_score
+                FROM lexical_scores
+            ),
+            scored AS (
+                SELECT
+                    slug,
+                    agent_name,
+                    agent_image,
+                    creator_username,
+                    creator_avatar,
+                    sub_heading,
+                    description,
+                    runs,
+                    rating,
+                    categories,
+                    featured,
+                    is_available,
+                    updated_at,
+                    semantic_score,
+                    lexical_score,
+                    category_score,
+                    recency_score,
+                    (
+                        {weights.lexical} * lexical_score +
+                        {weights.category} * category_score +
+                        {weights.recency} * recency_score
+                    ) as combined_score
+                FROM normalized
+            ),
+            filtered AS (
+                SELECT
+                    *,
+                    COUNT(*) OVER () as total_count
+                FROM scored
+                WHERE combined_score >= {min_score}
+            )
+            SELECT * FROM filtered
+            ORDER BY combined_score DESC
+            LIMIT ${param_index} OFFSET ${param_index + 1}
+        """
+
+        params.extend([page_size, offset])
+
+    try:
+        # Execute search query - includes total_count via window function
+        results = await query_raw_with_schema(sql_query, *params)
+
+        # Extract total count from first result (all rows have same count)
+        total = results[0]["total_count"] if results else 0
+
+        # Remove total_count from results before returning
+        for result in results:
+            result.pop("total_count", None)
+
+        logger.info(
+            f"Hybrid search for '{query}': {len(results)} results, {total} total "
+            f"(hybrid={use_hybrid})"
+        )
+
+        return results, total
+
+    except Exception as e:
+        logger.error(f"Hybrid search failed: {e}")
+        raise
+
+
+async def hybrid_search_simple(
+    query: str,
+    page: int = 1,
+    page_size: int = 20,
+) -> tuple[list[dict[str, Any]], int]:
+    """
+    Simplified hybrid search for common use cases.
+
+    Uses default weights and no filters.
+    """
+    return await hybrid_search(
+        query=query,
+        page=page,
+        page_size=page_size,
+    )

autogpt_platform/backend/migrations/20260109181714_add_docs_embedding/migration.sql

@@ -0,0 +1,35 @@
+-- CreateExtension in public schema (standard location for pgvector)
+CREATE EXTENSION IF NOT EXISTS "vector" WITH SCHEMA "public";
+
+-- Grant usage on public schema to platform users
+GRANT USAGE ON SCHEMA public TO postgres;
+
+-- CreateEnum
+CREATE TYPE "ContentType" AS ENUM ('STORE_AGENT', 'BLOCK', 'INTEGRATION', 'DOCUMENTATION', 'LIBRARY_AGENT');
+
+-- CreateTable
+CREATE TABLE "UnifiedContentEmbedding" (
+    "id" TEXT NOT NULL,
+    "createdAt" TIMESTAMP(3) NOT NULL DEFAULT CURRENT_TIMESTAMP,
+    "updatedAt" TIMESTAMP(3) NOT NULL,
+    "contentType" "ContentType" NOT NULL,
+    "contentId" TEXT NOT NULL,
+    "userId" TEXT,
+    "embedding" public.vector(1536) NOT NULL,
+    "searchableText" TEXT NOT NULL,
+    "metadata" JSONB NOT NULL DEFAULT '{}',
+
+    CONSTRAINT "UnifiedContentEmbedding_pkey" PRIMARY KEY ("id")
+);
+
+-- CreateIndex
+CREATE INDEX "UnifiedContentEmbedding_contentType_idx" ON "UnifiedContentEmbedding"("contentType");
+
+-- CreateIndex
+CREATE INDEX "UnifiedContentEmbedding_userId_idx" ON "UnifiedContentEmbedding"("userId");
+
+-- CreateIndex
+CREATE INDEX "UnifiedContentEmbedding_contentType_userId_idx" ON "UnifiedContentEmbedding"("contentType", "userId");
+
+-- CreateIndex
+CREATE UNIQUE INDEX "UnifiedContentEmbedding_contentType_contentId_userId_key" ON "UnifiedContentEmbedding"("contentType", "contentId", "userId");

autogpt_platform/backend/pyproject.toml

@@ -135,6 +135,9 @@ ignore_patterns = []
 [tool.pytest.ini_options]
 asyncio_mode = "auto"
 asyncio_default_fixture_loop_scope = "session"
+# Disable syrupy plugin to avoid conflict with pytest-snapshot
+# Both provide --snapshot-update argument causing ArgumentError
+addopts = "-p no:syrupy"
 filterwarnings = [
     "ignore:'audioop' is deprecated:DeprecationWarning:discord.player",
     "ignore:invalid escape sequence:DeprecationWarning:tweepy.api",

autogpt_platform/backend/schema.prisma

@@ -1,14 +1,15 @@
 datasource db {
-  provider  = "postgresql"
-  url       = env("DATABASE_URL")
-  directUrl = env("DIRECT_URL")
+  provider   = "postgresql"
+  url        = env("DATABASE_URL")
+  directUrl  = env("DIRECT_URL")
+  extensions = [pgvector(map: "vector")]
 }
 
 generator client {
   provider               = "prisma-client-py"
   recursive_type_depth   = -1
   interface              = "asyncio"
-  previewFeatures        = ["views", "fullTextSearch"]
+  previewFeatures        = ["views", "fullTextSearch", "postgresqlExtensions"]
   partial_type_generator = "backend/data/partial_types.py"
 }
 
@@ -127,8 +128,8 @@ model BuilderSearchHistory {
   updatedAt DateTime @default(now()) @updatedAt
 
   searchQuery String
-  filter      String[]  @default([])
-  byCreator   String[]  @default([])
+  filter      String[] @default([])
+  byCreator   String[] @default([])
 
   userId String
   User   User   @relation(fields: [userId], references: [id], onDelete: Cascade)
@@ -721,26 +722,25 @@ view StoreAgent {
   storeListingVersionId String
   updated_at            DateTime
 
-  slug                String
-  agent_name          String
-  agent_video         String?
-  agent_output_demo   String?
-  agent_image         String[]
+  slug              String
+  agent_name        String
+  agent_video       String?
+  agent_output_demo String?
+  agent_image       String[]
 
-  featured         Boolean                  @default(false)
-  creator_username String?
-  creator_avatar   String?
-  sub_heading      String
-  description      String
-  categories       String[]
-  search           Unsupported("tsvector")? @default(dbgenerated("''::tsvector"))
-  runs                Int
-  rating              Float
-  versions            String[]
-  agentGraphVersions  String[]
-  agentGraphId        String
-  is_available        Boolean                  @default(true)
-  useForOnboarding    Boolean                  @default(false)
+  featured           Boolean                  @default(false)
+  creator_username   String?
+  creator_avatar     String?
+  sub_heading        String
+  description        String
+  categories         String[]
+  runs               Int
+  rating             Float
+  versions           String[]
+  agentGraphVersions String[]
+  agentGraphId       String
+  is_available       Boolean                  @default(true)
+  useForOnboarding   Boolean                  @default(false)
 
   // Materialized views used (refreshed every 15 minutes via pg_cron):
   // - mv_agent_run_counts - Pre-aggregated agent execution counts by agentGraphId
@@ -856,14 +856,14 @@ model StoreListingVersion {
   AgentGraph        AgentGraph @relation(fields: [agentGraphId, agentGraphVersion], references: [id, version])
 
   // Content fields
-  name                String
-  subHeading          String
-  videoUrl            String?
-  agentOutputDemoUrl  String?
-  imageUrls           String[]
-  description         String
-  instructions        String?
-  categories          String[]
+  name               String
+  subHeading         String
+  videoUrl           String?
+  agentOutputDemoUrl String?
+  imageUrls          String[]
+  description        String
+  instructions       String?
+  categories         String[]
 
   isFeatured Boolean @default(false)
 
@@ -899,6 +899,9 @@ model StoreListingVersion {
   // Reviews for this specific version
   Reviews StoreListingReview[]
 
+  // Note: Embeddings now stored in UnifiedContentEmbedding table
+  // Use contentType=STORE_AGENT and contentId=storeListingVersionId
+
   @@unique([storeListingId, version])
   @@index([storeListingId, submissionStatus, isAvailable])
   @@index([submissionStatus])
@@ -906,6 +909,42 @@ model StoreListingVersion {
   @@index([agentGraphId, agentGraphVersion]) // Non-unique index for efficient lookups
 }
 
+
+// Content type enum for unified search across store agents, blocks, docs
+// Note: BLOCK/INTEGRATION are file-based (Python classes), not DB records
+// DOCUMENTATION are file-based (.md files), not DB records  
+// Only STORE_AGENT and LIBRARY_AGENT are stored in database
+enum ContentType {
+  STORE_AGENT     // Database: StoreListingVersion
+  BLOCK           // File-based: Python classes in /backend/blocks/
+  INTEGRATION     // File-based: Python classes (blocks with credentials)
+  DOCUMENTATION   // File-based: .md/.mdx files
+  LIBRARY_AGENT   // Database: User's personal agents
+}
+
+// Unified embeddings table for all searchable content types
+// Supports both public content (userId=null) and user-specific content (userId=userID)
+model UnifiedContentEmbedding {
+  id          String      @id @default(uuid())
+  createdAt   DateTime    @default(now())
+  updatedAt   DateTime    @updatedAt
+  
+  // Content identification
+  contentType ContentType
+  contentId   String      // DB ID (storeListingVersionId) or file identifier (block.id, file_path)
+  userId      String?     // NULL for public content (store, blocks, docs), userId for private content (library agents)
+  
+  // Search data  
+  embedding      Unsupported("public.vector(1536)")  // pgvector embedding from public schema
+  searchableText String                        // Combined text for search and fallback
+  metadata       Json        @default("{}")    // Content-specific metadata
+  
+  @@unique([contentType, contentId, userId])  // Allow same content for different users
+  @@index([contentType])
+  @@index([userId])
+  @@index([contentType, userId])
+}
+
 model StoreListingReview {
   id        String   @id @default(uuid())
   createdAt DateTime @default(now())
@@ -998,16 +1037,16 @@ model OAuthApplication {
   updatedAt DateTime @updatedAt
 
   // Application metadata
-  name               String
-  description        String?
-  logoUrl            String? // URL to app logo stored in GCS
-  clientId           String @unique
-  clientSecret       String // Hashed with Scrypt (same as API keys)
-  clientSecretSalt   String // Salt for Scrypt hashing
+  name             String
+  description      String?
+  logoUrl          String? // URL to app logo stored in GCS
+  clientId         String  @unique
+  clientSecret     String // Hashed with Scrypt (same as API keys)
+  clientSecretSalt String // Salt for Scrypt hashing
 
   // OAuth configuration
   redirectUris String[] // Allowed callback URLs
-  grantTypes   String[] @default(["authorization_code", "refresh_token"])
+  grantTypes   String[]           @default(["authorization_code", "refresh_token"])
   scopes       APIKeyPermission[] // Which permissions the app can request
 
   // Application management