Conversation
Introduces a 'lean' mode to _repo_search_impl that strips debug/internal fields from search results for improved token efficiency, controlled by a parameter or LEAN_RESPONSES env var. Refactors hybrid_search to support thread-safe, per-request ReFRAG config using contextvars and explicit parameters, avoiding global env var mutation. Updates TOON formatting to support lean mode and omits results_json when lean is enabled. Also improves Neo4j graph 'toon' output to include metadata for consistency.
Introduces a new Helm chart for deploying the context-engine application, including templates for deployments, services, ingress, configmaps, autoscaling, and persistent storage. This chart enables configurable, production-ready Kubernetes deployment of the context-engine and its components such as Qdrant, MCP indexer/memory, upload service, watcher, and learning reranker worker.
🤖 Augment PR SummarySummary: Adds Kubernetes deployment support via a new Helm chart, and refines MCP tool outputs to be more token-efficient. Changes:
Technical Notes: The chart relies heavily on ConfigMaps for runtime config and defaults to exposing several services via NodePort; MCP output size is now influenced by 🤖 Was this summary useful? React with 👍 or 👎 |
![augmentcode[bot]](https://avatars.githubusercontent.com/in/1027498?s=60&v=4){kind=small}
scripts/hybrid_search.py
Outdated
| Returns empty dict if context_answer module not available or not in request context. | ||
| """ | ||
| try: | ||
| from scripts.mcp_impl.context_answer import get_refrag_config |
There was a problem hiding this comment.
_get_contextvar_refrag_config() tries to import get_refrag_config from scripts.mcp_impl.context_answer, but that function doesn’t exist there, so this will always fall back to {}/env vars and won’t provide the per-request isolation described in the comment.
🤖 Was this useful? React with 👍 or 👎
|
|
||
| # Use injected lock or fall back to module-level lock | ||
| _lock = env_lock if env_lock is not None else _CA_ENV_LOCK | ||
| del env_lock # unused |
There was a problem hiding this comment.
env_lock is now ignored, but this function still mutates process-wide os.environ for REFRAG_*/MICRO_BUDGET_TOKENS later; concurrent requests can race and restore the wrong values for each other.
🤖 Was this useful? React with 👍 or 👎
|
|
||
| # Determine if lean mode is enabled (strips debug/internal fields for agent ROI) | ||
| # Default ON for better agent token efficiency; set LEAN_RESPONSES=0 to disable | ||
| _lean = _to_bool(lean, os.environ.get("LEAN_RESPONSES", "1").lower() not in ("0", "false", "no")) |
There was a problem hiding this comment.
Lean mode defaults to enabled (LEAN_RESPONSES defaulting to "1"), which changes the repo_search response schema (e.g., drops args/rerank_counters) and can break in-repo consumers/tests that read those fields.
🤖 Was this useful? React with 👍 or 👎
| if isinstance(results, list): | ||
| # Preserve original list for internal callers before TOON encoding | ||
| response["results_json"] = results | ||
| # Only preserve results_json if not in lean mode (saves tokens for agents) |
There was a problem hiding this comment.
In lean mode this skips emitting results_json, but several internal callers rely on results_json to consume TOON responses without decoding; those paths may break when TOON_ENABLED=1 and lean is enabled.
🤖 Was this useful? React with 👍 or 👎
| return _format_neo4j_graph_toon(response) | ||
| return { | ||
| "ok": True, | ||
| "result": _format_neo4j_graph_toon(response), |
There was a problem hiding this comment.
For output_format == "toon" the payload uses result (singular) instead of results, which is inconsistent with other tools and can cause clients that look for results to treat the response as empty.
🤖 Was this useful? React with 👍 or 👎
Introduces scripts/analyze_intent_confidence.py for analyzing intent classification confidence from event logs, along with comprehensive tests in tests/test_analyze_intent_confidence.py. Updates repo_search and related functions to support a 'lean' argument for improved internal composition and testability. Increases INDEX_UPSERT_BATCH and CPU allocation in docker-compose.yml for better indexing performance. Refactors tests and context handling for improved reliability and clarity.
1 participant
Helm charts., Lean agent output