Get update from llm-350

constants.ini

@@ -12,4 +12,5 @@ DOMAIN=localhost
 DB_PASSWORD=dbadmin
 RAG_SEARCH_RUUTER_PUBLIC_INTERNAL_SERVICE=http://ruuter:8086/services
 SERVICE_DMAPPER_HBS=http://data-mapper:3000/hbs/rag-search
-SERVICE_PROJECT_LAYER=services
+SERVICE_PROJECT_LAYER=services
+RAG_SEARCH_LLM_SERVICE=http://llm-orchestration-service:8100

src/api_tool_indexer/constants.py

@@ -30,8 +30,19 @@ class ApiToolIndexerConstants:
     RETRY_DELAY_BASE = 2  # Exponential backoff base (2^attempt seconds)
     REQUEST_TIMEOUT = 60  # seconds
 
+    # Number of example queries generated per endpoint.
+    # Each example becomes its own Qdrant point so its vector sits in the exact
+    # language region of the embedding space, enabling short-query matching.
+    EXAMPLE_QUERY_COUNT = 5
+
     # Context Enrichment Template
-    # Used to generate a rich semantic context for each endpoint before embedding
+    # Full template goes in chunk_prompt; document_prompt is left empty.
+    #
+    # Multi-point indexing strategy:
+    #   - Each example query line is extracted and stored as its own Qdrant point,
+    #     embedded from that individual sentence alone.
+    #   - The prose + all examples combined become one summary point.
+    # All in the same language as the endpoint description — no bilingual duplication.
     CONTEXT_TEMPLATE = """<document>
 {full_endpoint_info}
 </document>
@@ -44,12 +55,23 @@ class ApiToolIndexerConstants:
 </endpoint>
 
 Please generate a rich, detailed context that describes this API endpoint comprehensively for semantic search.
-Include information about:
+Keep the prose context general and country-agnostic. Include information about:
 - What the user wants to accomplish by calling this endpoint
 - Key terms and synonyms for this action
 - Related concepts and use cases
 - Common ways users might ask for this functionality in natural language
 
-IMPORTANT: Generate the context in the SAME LANGUAGE as the endpoint description above. If the description is in Estonian, respond in Estonian. If in English, respond in English. If in Russian, respond in Russian.
+IMPORTANT: Generate the prose context and the example questions in the SAME LANGUAGE as the endpoint description above. However, always use the exact section header "Example queries:" in English regardless of language — this is a required machine-readable marker.
+
+IMPORTANT for example queries: This is a system built for Estonian government digital services (Bürokratt). Ground the examples in an Estonian context — use Estonian cities (Tallinn, Tartu, Pärnu, Narva), Estonian institutions, and Estonia-relevant scenarios. Only use non-Estonian locations if the endpoint is explicitly about comparing or fetching data for multiple countries.
+
+Then add a section with exactly {example_count} realistic and diverse example questions a real user might ask when they need this endpoint. Cover different phrasings, synonyms, and indirect ways of asking — do not just repeat the description verbatim.
+
+Example queries:
+- <example question 1>
+- <example question 2>
+- <example question 3>
+- <example question 4>
+- <example question 5>
 
-Answer only with the enriched context and nothing else."""
+Answer only with the enriched context and example queries — nothing else."""

src/api_tool_indexer/main_indexer.py

@@ -2,19 +2,27 @@
 
 Receives raw API EndpointData, enriches it with LLM-generated context,
 creates hybrid embeddings (dense + sparse), and stores the result in Qdrant
-api_tool_collection as a single point per endpoint.
+api_tool_collection as multiple points per endpoint.
+
+Multi-point indexing strategy:
+    - One 'example' point per example query extracted from the LLM context.
+      Each query is embedded individually so its vector sits in the correct
+      language region of the embedding space, enabling accurate short-query matching.
+    - One 'summary' point containing the combined name + description + enriched context.
+      This handles broad/paraphrased queries that don't match any single example.
 
 Pipeline steps:
     1. Build LLM prompt from endpoint name, description, and params
     2. Generate context via LLMAPIClient.generate_context()
-    3. Build embed text: name + description + context + param descriptions
-    4. Create dense embedding via LLMAPIClient.create_embedding()
-    5. Create sparse vector via compute_sparse_vector()
-    6. Delete existing Qdrant point for idempotent update
-    7. Upsert EnrichedEndpoint to api_tool_collection
+    3. Parse example query lines from the returned context
+    4. Create dense + sparse embeddings per example query (example points)
+    5. Create dense + sparse embedding for combined summary text (summary point)
+    6. Delete all existing Qdrant points for this endpoint (filter-based, idempotent)
+    7. Upsert all points to api_tool_collection
     8. Return IndexingResult
 """
 
+import re
 import sys
 import json
 import asyncio
@@ -111,10 +119,13 @@ async def _generate_context_for_endpoint(
 
     logger.info(f"params_summary : {params_summary}")
 
+    # Escape braces in the URL to prevent str.format() from treating path
+    # parameter templates like {id} as format placeholders (KeyError).
+    safe_url = endpoint_data.url.replace("{", "{{").replace("}", "}}")
     full_endpoint_info = (
         f"Endpoint: {endpoint_data.name}\n"
         f"Method: {endpoint_data.method}\n"
-        f"URL: {endpoint_data.url}\n"
+        f"URL: {safe_url}\n"
         f"Description: {endpoint_data.description}\n"
         f"Parameters: {params_summary}"
     )
@@ -124,15 +135,25 @@ async def _generate_context_for_endpoint(
         name=endpoint_data.name,
         description=endpoint_data.description,
         params_summary=params_summary,
+        example_count=ApiToolIndexerConstants.EXAMPLE_QUERY_COUNT,
+    )
+
+    logger.debug(
+        "Generated context prompt for endpoint '{}': {} chars",
+        endpoint_data.endpoint_id,
+        len(context_prompt),
     )
 
-    # Re-use the internal HTTP call of LLMAPIClient - /generate-context endpoint
+    # context_type="api_tool" makes context_manager use API_TOOL_CONTEXT_PROMPT,
+    # which passes chunk_prompt through unmodified so CHUNK_CONTEXT_PROMPT cannot
+    # override the instructions in CONTEXT_TEMPLATE (e.g. example query generation).
     request_data = {
         "document_prompt": "",
         "chunk_prompt": context_prompt,
         "environment": api_client.environment,
-        "use_cache": True,
+        "use_cache": False,
         "connection_id": api_client.connection_id,
+        "context_type": "api_tool",
     }
 
     last_error = None
@@ -153,6 +174,13 @@ async def _generate_context_for_endpoint(
             result = response.json()
 
             context = result.get("context", "").strip()
+
+            logger.debug(
+                "context preview: {}{}",
+                context[:200].replace("\n", "\\n"),
+                "..." if len(context) > 200 else "",
+            )
+
             if not context:
                 raise ValueError("Empty context returned from API")
 
@@ -181,9 +209,54 @@ async def _generate_context_for_endpoint(
     raise RuntimeError(error_msg)
 
 
+_EXAMPLE_SECTION_HEADER = re.compile(r"^example queries\s*:", re.IGNORECASE)
+
+
+def _parse_example_queries(context: str) -> List[str]:
+    """Extract example query lines from the LLM-generated context.
+
+    Scans for the 'Example queries:' section header and collects every
+    subsequent '- ' line until the section ends.
+
+    Args:
+        context: Raw LLM-generated context string from generate_context().
+
+    Returns:
+        List of example query strings, deduplicated and preserving order.
+    """
+    examples: List[str] = []
+    in_section = False
+
+    for line in context.splitlines():
+        stripped = line.strip()
+        if _EXAMPLE_SECTION_HEADER.match(stripped):
+            in_section = True
+            continue
+        if in_section:
+            if stripped.startswith("- "):
+                examples.append(stripped[2:].strip())
+            elif stripped and not stripped.startswith("#"):
+                # Non-empty, non-comment line that isn't a list item ends the section
+                in_section = False
+
+    # Deduplicate preserving order
+    seen: set[str] = set()
+    unique: List[str] = []
+    for ex in examples:
+        if ex and ex not in seen:
+            seen.add(ex)
+            unique.append(ex)
+    return unique
+
+
 async def index_endpoint(endpoint_data: EndpointData) -> IndexingResult:
     """Index one API endpoint into Qdrant api_tool_collection.
 
+    Creates multiple points per endpoint:
+    - One 'example' point per parsed example query, embedded from that
+      individual sentence so the vector sits in the correct language region.
+    - One 'summary' point embedded from the combined name + description + context.
+
     Args:
         endpoint_data: Raw endpoint data from mock_endpoints table.
 
@@ -197,7 +270,6 @@ async def index_endpoint(endpoint_data: EndpointData) -> IndexingResult:
     )
 
     try:
-        # Steps 1–5: LLM enrichment and embedding
         async with LLMAPIClient(
             api_base_url=ApiToolIndexerConstants.DEFAULT_API_BASE_URL,
             environment=ApiToolIndexerConstants.DEFAULT_ENVIRONMENT,
@@ -206,92 +278,131 @@ async def index_endpoint(endpoint_data: EndpointData) -> IndexingResult:
             retry_delay_base=ApiToolIndexerConstants.RETRY_DELAY_BASE,
             timeout=ApiToolIndexerConstants.REQUEST_TIMEOUT,
         ) as api_client:
-            # Step 1-2: Generate LLM enriched context
-            logger.info("Step 1/5: Generating LLM enriched context")
+            # Step 1: Generate LLM enriched context (prose + example queries)
+            logger.info("Step 1/4: Generating LLM enriched context")
             enriched_context = await _generate_context_for_endpoint(
                 api_client, endpoint_data
             )
 
-            # Step 3: Build embed text combining all semantic signal
+            # Step 2: Parse example query lines from the context
+            example_queries = _parse_example_queries(enriched_context)
+            if not example_queries:
+                logger.warning(
+                    f"No example queries parsed from context for endpoint '{endpoint_id}'. "
+                    "The LLM output may not contain an 'Example queries:' section. "
+                    "Only a summary point will be indexed — search accuracy may be reduced."
+                )
+            else:
+                logger.info(
+                    f"Step 2/4: Parsed {len(example_queries)} example queries from context"
+                )
+
+            # Step 3: Embed each example query individually → example points
+            logger.info(
+                f"Step 3/4: Creating embeddings for {len(example_queries)} example points"
+            )
+            enriched_points: List[EnrichedEndpoint] = []
+
+            for i, example in enumerate(example_queries):
+                logger.debug(
+                    f"  Embedding example {i + 1}/{len(example_queries)}: "
+                    f"'{example[:80]}{'...' if len(example) > 80 else ''}'"
+                )
+                ex_embedding = await api_client.create_embedding(example)
+                ex_sparse = compute_sparse_vector(example)
+                enriched_points.append(
+                    EnrichedEndpoint(
+                        endpoint_id=endpoint_id,
+                        name=endpoint_data.name,
+                        description=endpoint_data.description,
+                        url=endpoint_data.url,
+                        method=endpoint_data.method,
+                        params=endpoint_data.params,
+                        enriched_context=enriched_context,
+                        service_id=endpoint_data.service_id,
+                        point_type="example",
+                        example_text=example,
+                        embedding=ex_embedding,
+                        sparse_indices=ex_sparse.indices,
+                        sparse_values=ex_sparse.values,
+                    )
+                )
+
+            # Step 4: Embed combined summary text → summary point
+            logger.info("Step 4/4: Creating summary point embedding")
             params_summary = _build_params_summary(endpoint_data.params)
-            embed_text = (
+            summary_text = (
                 f"{endpoint_data.name}. "
                 f"{endpoint_data.description}. "
                 f"{enriched_context}. "
                 f"Parameters: {params_summary}"
             )
+            summary_embedding = await api_client.create_embedding(summary_text)
 
-            # Step 4: Create dense embedding vector
-            logger.info("Step 2/5: Creating dense embedding vector")
-            dense_embedding = await api_client.create_embedding(embed_text)
-
-        # Step 5: Create sparse (BM25) vector - synchronous, after closing HTTP session
-        logger.info("Step 3/5: Computing sparse (BM25) vector")
-        sparse_vec = compute_sparse_vector(embed_text)
-
-        # Build EnrichedEndpoint ready for Qdrant storage
-        enriched = EnrichedEndpoint(
-            endpoint_id=endpoint_id,
-            name=endpoint_data.name,
-            description=endpoint_data.description,
-            url=endpoint_data.url,
-            method=endpoint_data.method,
-            params=endpoint_data.params,
-            enriched_context=enriched_context,
-            service_id=endpoint_data.service_id,
-            embedding=dense_embedding,
-            sparse_indices=sparse_vec.indices,
-            sparse_values=sparse_vec.values,
+        # Sparse vectors are CPU-bound — computed after the HTTP session closes
+        summary_sparse = compute_sparse_vector(summary_text)
+        enriched_points.append(
+            EnrichedEndpoint(
+                endpoint_id=endpoint_id,
+                name=endpoint_data.name,
+                description=endpoint_data.description,
+                url=endpoint_data.url,
+                method=endpoint_data.method,
+                params=endpoint_data.params,
+                enriched_context=enriched_context,
+                service_id=endpoint_data.service_id,
+                point_type="summary",
+                embedding=summary_embedding,
+                sparse_indices=summary_sparse.indices,
+                sparse_values=summary_sparse.values,
+            )
         )
 
-        # Steps 6-7: Qdrant operations (separate try/finally ensures connection is closed)
+        # Qdrant operations — separate block so the connection is always closed
         qdrant = ApiToolQdrantManager()
         try:
             qdrant.connect()
             qdrant.ensure_collection()
 
-            # Step 6: Delete existing point for idempotent update
-            logger.info("Step 4/5: Deleting existing Qdrant point (idempotent update)")
-            deleted = qdrant.delete_endpoint_point(endpoint_id)
+            # Delete all existing points for this endpoint (filter-based, idempotent)
+            deleted = qdrant.delete_endpoint_points(endpoint_id)
             if not deleted:
                 logger.error(
-                    f"Failed to delete existing Qdrant point for endpoint '{endpoint_id}'. "
+                    f"Failed to delete existing points for endpoint '{endpoint_id}'. "
                     "Aborting upsert to prevent stale data."
                 )
                 return IndexingResult(
                     success=False,
                     endpoint_id=endpoint_id,
                     message="Qdrant delete failed before upsert",
-                    error="delete_endpoint_point returned False",
+                    error="delete_endpoint_points returned False",
                 )
 
-            # Step 7: Upsert the enriched endpoint
-            logger.info("Step 5/5: Upserting endpoint into api_tool_collection")
-            upserted = qdrant.upsert_endpoint(enriched)
-
+            upserted = qdrant.upsert_endpoint_points(enriched_points)
         finally:
             qdrant.close()
 
-        # Step 8: Return result
+        n_examples = len(example_queries)
         if upserted:
             logger.success(
-                f"Endpoint '{endpoint_id}' (name='{endpoint_data.name}') "
-                "indexed successfully"
+                f"Endpoint '{endpoint_id}' (name='{endpoint_data.name}') indexed successfully "
+                f"({n_examples} example points + 1 summary point)"
             )
             return IndexingResult(
                 success=True,
                 endpoint_id=endpoint_id,
                 message=(
                     f"Endpoint '{endpoint_data.name}' indexed successfully into "
-                    f"api_tool_collection (dim={len(dense_embedding)})"
+                    f"api_tool_collection "
+                    f"({n_examples} example points + 1 summary point)"
                 ),
             )
         else:
             return IndexingResult(
                 success=False,
                 endpoint_id=endpoint_id,
                 message="Qdrant upsert failed",
-                error="upsert_endpoint returned False",
+                error="upsert_endpoint_points returned False",
             )
 
     except Exception as e: