fix: contextual retrieval latency — max_chunks gate + prompt trim + model update

feat: chat route accepts filter_docs for scoped retrieval

feat: add filter_docs to HybridRetriever and get_retriever_filtered helper

fix: revert out-of-scope model default change in contextualize_chunks

feat: add source_type metadata to all ingested chunks

fix: restore score > 0 threshold in filtered BM25 path + use original index for scoring

feat: add filter_sources param to BM25Index.search()

fix: citation toggle stopPropagation + absolute UPLOAD_DIR path

fix: replace startswith with is_relative_to() for path traversal security

feat: add GET /api/files/{filename} route for citation PDF links

feat: HyDE re-eval (recall +21pp), briefing JSON fix, eval run cleanup

fix: remove web_search field, ensure cleanup runs on all paths in chat route

feat: stream POST /api/chat via SSE StreamingResponse

feat: add stream_query_with_web async generator to chain.py

feat: semantic chunking via SemanticChunker, toggled by config.yaml semantic_enabled

fix: strip trailing commas from LLM JSON in briefing generation

feat: async embed upload — return 202 immediately, poll status for stage labels

refactor: remove Render 512MB memory guards — running on HF Spaces 16GB

fix: RSS guard in contextual refresh — skip if base RSS >460MB

fix: reduce peak RSS during upload — reorder GC + briefing after embed

fix: OOM guard — skip web search during contextual refresh + psutil memory logging

fix: reduce max_concurrent 20→3 to stay under Groq 6k TPM; smart 429 retry wait

fix: return 422 on password-protected PDF upload instead of 500

feat: contextual retrieval in production via async BackgroundTask

feat: add contextualize_chunks() for LLM-prefixed chunk embeddings

feat: multi-query retrieval + HyDE/MQ eval runs (v1.1.0, v1.2.0 Violet)

feat: Stage 8 — eval dashboard, mandatory web search, docs update

feat: add HyDE (Hypothetical Document Embeddings) for dense retrieval

fix: singleton vectorstore/retriever cache to prevent OOM on repeated queries

fix: URL size guard for OOM prevention, eliminate double retrieval in chat, fix print

fix: workspace-aware retrieval in chat handler, robust list_collections() for chromadb 0.5+

feat: multi-workspace backend — workspace_id routes ChromaDB collections + BM25 dict

fix: use langchain_core.prompts for PromptTemplate, propagate citation_index in run_query

feat: add citation_index to retriever + numbered doc prompt in QA chain

fix: robust JSON extraction in briefing (handles LLM preamble), tighten tests

feat: add briefing module (5-bullet summary + 3 suggested Qs on upload)

fix: sanitize upload_url error messages, handle embed failure, add timeout

feat: add URL ingestion endpoint + file/url tab UI in FileUpload

fix: use httpx.HTTPStatusError in url_loader test, widen timeout type to float

feat: add URL loader with BeautifulSoup extraction + tests

feat: rebrand FinRAG → Prism (collection names, system prompt, UI branding)

refactor: Remove live RAGAS eval, replace with static benchmark panel

fix: Graceful RAGAS 500 + info tooltips on eval metrics and faithfulness badge

fix: Switch reranker to TinyBERT-L-2-v2 (~17MB) to prevent OOM on web queries

fix: Reduce web search memory footprint to prevent OOM on second query

fix: Reduce CrossEncoder memory spike to prevent OOM on Render 512MB

fix: Lazy-import ragas to prevent startup crash on Render

feat: Migrate LLM to Groq, fix web search pipeline, add structured logging

feat: Add RAGAS eval, score passthrough, docs, and hybrid retrieval config

feat: Add hybrid retrieval with BM25, RRF fusion, and cross-encoder reranker