search-layer

Run the search.py script with a query and mode flag to query Exa, Tavily, and Grok in parallel, while the agent adds Brave results via web_search. Results are merged, deduplicated by canonical URL, and scored by intent-weighted relevance.

Intent-aware ranking classifies a query into types like factual, news, or comparison, then adjusts scoring weights for keyword match, freshness, and domain authority. News queries weight freshness at 0.6 while factual queries weight authority at 0.5.

Use fetch_thread.py with the issue URL to retrieve the body, all comments, labels, and timeline events via the GitHub API. It also extracts structured references such as linked issues, PRs, commits, and duplicate markers.

Yes, the pipeline degrades gracefully per source. Missing Grok credentials fall back to Exa plus Tavily, Exa errors continue with the remaining sources, and Brave web_search alone always works as the final fallback.

Different providers index and rank content differently, so contradictions occur especially on fast-moving topics. The synthesis phase explicitly flags conflicting claims and attributes single-source or stale information to its origin.

Avoid chain tracking for simple factual or resource lookups where a single answer suffices, since recursive fetching with LLM scoring adds latency and API cost. It is designed for status and exploratory queries needing deep thread analysis.