What's Changed
PR: #856
-
New Features
- Groq-backed completion, query expansion, and query interpretation.
- LLM reranking via Cohere, Groq, or OpenAI with priority Cohere → Groq → OpenAI.
- Dynamic operation selection based on available API keys.
- New config keys: GROQ_API_KEY, COHERE_API_KEY, CEREBRAS_API_KEY.
- Higher fetch limits for reranking (backend up to 250; frontend default 100).
-
Refactors
- Removed streaming delta events for expansion/interpretation; only start/done events emit.
- Fail-fast validation: non-empty query, strict Qdrant filter normalization, reranking requires API key, interpretation is required when enabled.
- Qdrant search enforces a top-level limit; decay config validation improved.
- Recency bias skips gracefully if destination creation fails; logging moved to debug where noisy.
- Dependencies updated: openai 1.109.1; added groq and cohere.