fix(query): resolve contradictory retry instruction and review feedback

Address claudebot review on PR #37:
- Critical: GENERATE_PROMPT instruction #0 was still telling the model to "switch to let+pipeline form" on retry — leftover from before we discovered Cosmos rejects let/pipeline inside $lookup. Both CROSS_COLLECTION_GUIDANCE and EVALUATE_PROMPT already ban that form, so the retry loop would cycle. Rewrite to recommend the $addFields + $toObjectId/$toString pre-conversion pattern.
- Medium: Replace the magic string "Could not fetch schema summary." with a SCHEMA_FETCH_FAILED module-level sentinel so routes/query.py is no longer coupled to an exact error string.
- Low: Use logging.getLogger(__name__).warning(...) in routes/query.py instead of print() for the relationship-inference failure path.
- CI: black formatting (re-run on react_agent_service.py).

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Author: ChingEnLin

backend/routes/query.py

@@ -23,6 +23,7 @@
     execute_mongo_query,
     transform_mongo_result,
     get_database_schema_summary,
+    SCHEMA_FETCH_FAILED,
 )
 from models.analyze import AnalyzeRequest, AnalyzeResponse
 from services.analyze_service import analyze_query_result
@@ -36,9 +37,12 @@
     DeleteResult,
 )
 import ast
+import logging
 import re
 from google import genai
 
+logger = logging.getLogger(__name__)
+
 router = APIRouter()
 
 
@@ -105,7 +109,7 @@ def nl2query(prompt: QueryPrompt = Body(...), authorization: str = Header(...)):
     if (
         len(collections) > 1
         and schema_summary
-        and schema_summary != "Could not fetch schema summary."
+        and schema_summary != SCHEMA_FETCH_FAILED
     ):
         try:
             rels = generate_schema_relationships(schema_summary, model=prompt.model)
@@ -117,7 +121,9 @@ def nl2query(prompt: QueryPrompt = Body(...), authorization: str = Header(...)):
                     for r in rels.relationships
                 )
         except Exception as e:
-            print(f"[WARN] Relationship inference failed; continuing without it: {e}")
+            logger.warning(
+                "Relationship inference failed; continuing without it: %s", e
+            )
 
     return run_query_generator(
         user_input=prompt.user_input,

backend/services/mongo_service.py

@@ -1,4 +1,6 @@
 from bson import ObjectId
+
+SCHEMA_FETCH_FAILED = "Could not fetch schema summary."
 import pymongo
 from pymongo.results import (
     UpdateResult,
@@ -90,4 +92,4 @@ def get_database_schema_summary(
         return "\n\n".join(summary)
     except Exception as e:
         print(f"Error fetching schema summary: {e}")
-        return "Could not fetch schema summary."
+        return SCHEMA_FETCH_FAILED

backend/services/react_agent_service.py

@@ -151,7 +151,7 @@ class AgentState(TypedDict):
 {evaluation}
 
 Instructions:
-0. If a Previous Attempt is shown above, your new query MUST be materially different from it — change the join form, fields, types, or stages in response to the critique. Producing the same query (or a trivial reformat) is a failure. In particular, if the previous attempt used `localField`/`foreignField` $lookup and returned empty for every doc, switch to the `let` + `pipeline` form with $toObjectId/$toString to fix likely type mismatches.
+0. If a Previous Attempt is shown above, your new query MUST be materially different — change the join form, fields, types, or stages in response to the critique. Producing the same query (or a trivial reformat) is a failure. If the previous $lookup returned empty arrays for every input doc, the cause is almost always a type mismatch on the join key: add an `$addFields` stage BEFORE the `$lookup` that applies `$toObjectId` (string→OID) or `$toString` (OID→string), then $lookup on the converted field. NEVER use `let` or `pipeline` inside `$lookup` — Cosmos rejects both with CommandNotSupported.
 1. Write ONLY the PyMongo query code.
 2. Use variables appropriately (e.g., db['collection_name'].find(...) or db['collection_name'].aggregate(...)).
 3. Do not include markdown formatting or explanations, just return the code, but you can use ```python if you must.
@@ -199,11 +199,14 @@ def generate_query_node(state: AgentState):
         database=state["database"],
         collections=", ".join(state["collections"]),
         schema_context=state["schema_context"],
-        relationship_context=state.get("relationship_context") or "None (single collection or no inferred relationships).",
+        relationship_context=state.get("relationship_context")
+        or "None (single collection or no inferred relationships).",
         intermediate_context=state.get("intermediate_context", {}),
         previous_query=previous_query or "None (this is the first attempt).",
         evaluation=state.get("evaluation", "None"),
-        cross_collection_guidance=CROSS_COLLECTION_GUIDANCE if is_multi_collection else "",
+        cross_collection_guidance=(
+            CROSS_COLLECTION_GUIDANCE if is_multi_collection else ""
+        ),
     )
 
     try: