fix(core): improve retrieval quality and pipeline observability#52

Merged

fvadicamo merged 33 commits into

fix/conversation-persistence

Mar 25, 2026

Commits on Mar 24, 2026

fix(core): register conversation store in provider registry

fvadicamo
and
claude
committed
chore(backlog): add BUG-014 conversation persistence gap

fvadicamo
and
claude
committed
docs(claude): add API interaction directive

fvadicamo
and
claude
committed
fix(core): create conversation row before first turn (BUG-014)

fvadicamo
and
claude
committed
fix(core): prevent LLM from exposing RAG internals in any language

fvadicamo
and
claude
committed
fix(core): tighten system prompt against implicit RAG exposure

fvadicamo
and
claude
committed
fix(core): simplify system prompt to improve LLM compliance

fvadicamo
and
claude
committed
fix(core): rewrite system prompt following RAG best practices research

fvadicamo
and
claude
committed
fix(core): wire VEKTRA_LLM_API_KEY to litellm and add extra_body support

fvadicamo
and
claude
committed
refactor(learn): clean up conversation persistence code

fvadicamo
and
claude
committed
chore(backlog): add BUG-015, BUG-016, TECH-002, DEBT-010 for retrieval quality

fvadicamo
and
claude
committed
fix(build): harden supply chain - eliminate uv pip install outside lockfile

fvadicamo
and
claude
committed
fix(core): propagate reranker scores to SearchResult (BUG-015)

fvadicamo
and
claude
committed
feat(core): switch to multilingual reranker and lower threshold (BUG-016, DEBT-010)

fvadicamo
and
claude
committed
feat(eval): add retrieval and e2e evaluation harness (TECH-002)

fvadicamo
and
claude
committed
fix(eval): add --use-query flag to eval_retrieval for reranker evaluation

fvadicamo
and
claude
committed

Commits on Mar 25, 2026