About TB Research

tbresearch.orgis the open, machine-readable archive of the world's tuberculosis knowledge. We pull every published TB paper from the open biomedical literature, normalize it to markdown, embed it, and serve hybrid keyword + semantic search with cited RAG answers. The full vision is in docs/02_prd.md.

Current corpus

• 154661 documents · 638246 embedded chunks · publication years 1944–2026
• Sources: PubMed (E-utilities), Europe PMC, ClinicalTrials.gov, OpenAlex citation graph
• Embeddings: nomic-embed-text (768d) via self-hosted Ollama
• Answer model: Claude Sonnet 4.6 with strict per-claim citation requirements (PRD §10.3)

What you can do today

• Search the corpus with hybrid BM25 + cosine similarity, fused via reciprocal rank fusion. Filter by document type, TB subtype, year, country, or language.
• Ask TB Research a research question and get a synthesized answer with inline [tbrsch_id]citations to specific documents. The model is instructed to say “I don't know” rather than fabricate, and to refuse clinical decision support for specific patients.
• Browse a document at /doc/[tbrsch_id] with full metadata, related documents, and citation graph.
• Subscribe to a saved search via RSS for new-paper alerts.
• Use the public API at /v1/* for programmatic access: search, full document records, citation graph, and authenticated bulk export.

What we don't do yet

• Grey literature, NTP guidelines, conference abstracts, survivor narratives
• Translation pipeline for non-English documents
• Patient narrative consent & right-to-withdraw workflow
• Fine-tuned TB Research SLM (Phase 2)
• Clinician / patient / advocate UI surfaces (Phase 2 per PRD §11.2)

Architecture

[ PubMed · Europe PMC · ClinicalTrials.gov · OpenAlex ]
                       ↓
              [ ingest + normalize ]
                       ↓
          [ Postgres + pgvector + tsvector ]
                       ↓
                 [ hybridSearch ]
                       ↓
           ┌───────────┴───────────┐
           ↓                       ↓
     Search results       Ask TB Research (RAG)
                                   ↓
                       Claude Sonnet 4.6
                       + strict citations

Stack

• Next.js 16 on Vercel
• Supabase Postgres + pgvector + tsvector (IVFFlat index)
• Self-hosted Ollama on Fly.io for embeddings
• Anthropic Claude (Sonnet 4.6 answer, Haiku 4.5 rewrite/rerank)

Licensing

Code: Apache 2.0. Corpus metadata: CC0. Document content preserves its source license. Bulk exports surface license information via the X-TBRsch-License header so downstream consumers can honor it.