Spaces:
Sleeping
Sleeping
Commit History
-
docs: implementation plan for search latency optimization
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>b3f12dd -
docs: spec for search latency optimization (score gap + LLM fallback)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>1094935
-
docs: implementation plan for eval CLI
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>b145b51 -
docs: spec for eval CLI to iterate on retrieval quality
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>72296b0 -
docs: search interface implementation plan
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>475c2dd -
docs: search interface design spec
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>5608593 -
docs: mosaic redesign implementation plan
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>5293fe8 -
docs: mosaic redesign spec — masonry, infinite scroll, randomization
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>eb7f6f5
-
docs: custom frontend design spec and implementation plan
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>6e73893 -
docs: implementation plan for ONNX lightweight migration
6 tasks: export model, rewrite embedder, remove reranker, update entrypoints, update build, re-vectorize. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>043a5bb -
docs: design spec for ONNX lightweight migration
Replace PyTorch + sentence-transformers (~2GB) with ONNX Runtime direct loading (~50MB, already a chromadb dep). Switch bge-large to bge-small, drop cross-encoder reranker. Target: HF Spaces deployable image ~400MB. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>bdb27e7
-
plan: hybrid RAG + enrichment implementation (16 tasks)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>6aca361 -
spec: corrections review AI engineer (RAG best practices)
- Eval: 50 queries minimum (5+ par intent), ajout Precision@5 + router accuracy - Enrichment: is_subject instruction explicite dans le prompt LLM - Reranker: texte condensé inclut summary tronqué (300 chars) - Post-filter: top_k=100 quand filtres non-scalaires actifs (risque recall) - Prompts: citation format story-only, plus de Speaker/Timestamp - Router: Phase 1 regex-only, LLM fallback différé à Phase 2 - Award intent bm25_boost 0.7 → 0.85 - Risque: memory footprint HF Spaces (BGE-large + cross-encoder) - published_year dérivé dans pipeline.py (int pour ChromaDB where) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>d732c00 -
spec: intégrer les corrections de la review indépendante
Corrections critiques : - Post-filtering Python pour metadata comma-separated (ChromaDB $contains non supporté) - Fallback safe quand le router LLM échoue - Router regex-first, LLM-fallback (élimine 300ms pour 80% des queries) - Cross-encoder L-12 sur texte condensé (limite 512 tokens) - BM25 tokenization avec stopwords + stemming (PyStemmer) - Filters storage-agnostic (découplé de ChromaDB) - PersonMention.is_subject remplace key_subjects (évite redondance) - Ajout theme "social impact & advocacy" (15 catégories) - Section évaluation (ground truth, MRR@5, Recall@5) - Section risques identifiés + mitigations - RRF k=20 (petit corpus) - Startup warmup des modèles Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>a65d96c -
spec: hybrid RAG + enrichissement élargi design
Architecture B retenue : hybrid search (dense + BM25) + cross-encoder reranking + query routing. Enrichissement élargi avec awards, themes (taxonomie fermée), countries, media_type, event_period parsés depuis les JSON plateforme + Gemini Flash. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>f357a16
-
Add implementation plan for poster images in RAG pipeline
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>0e84ba8 -
Clarify Gradio rendering and cache compatibility in spec
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>03d33f1 -
Add design spec for poster images in RAG pipeline
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>bcdca21