CLAUDE.md

Candle-Fire — Architecture Guide

What This Is

Candle-fire is a physician-facing ALS research intelligence tool. A physician asks a free-text question ("What's the evidence for tofersen targeting SOD1?") and gets a synthesized, cited answer grounded in ~500 curated ALS papers, enriched by a knowledge graph.

Sibling project: beacon (patient-facing clinical trial finder at ../beacon). Follow the same conventions.

Two-Layer Intelligence

1. Knowledge Graph (KG): NetworkX DiGraph linking Gene → Protein → Compound → Pathway → Phenotype → ClinicalTrial. Used to expand query entities before retrieval (e.g., "tofersen" → SOD1 → oxidative stress → related compounds).

2. RAG (Vector Search): ChromaDB collection of ~500 ALS paper abstracts/full-text. Citation-count-weighted re-ranking. Used to retrieve evidence passages for synthesis.

Query pipeline: KG expansion first, then RAG retrieval with expanded entity context, then Claude synthesis.

Module Responsibilities

File/Dir	Responsibility
config.py	All constants: model names, file paths, ALS seed entities, API endpoints
models.py	Dataclasses: ALSPaper, ExtractedEntity, EntityRelationship, ResearchLandscape
prompts.py	System prompts for extraction agent and synthesis agent
tools.py	Tool schema loader — reads JSON from data/tools/, exports typed tool params
llm.py	LLM provider abstraction (Anthropic/OpenAI switchable via LLM_PROVIDER env var)
logging_config.py	Structured JSON rotating log to logs/candle_fire.log
ingestion/pubmed.py	PubMed Entrez client: fetch abstracts + metadata by MeSH query or PMID list
ingestion/pmc.py	PMC XML full-text fetcher: structured section text for Open Access papers
ingestion/clinicaltrials.py	ClinicalTrials.gov v2 client for ALS trials (no geo/distance, unlike beacon)
ingestion/semantic_scholar.py	Citation count enrichment per PMID via Semantic Scholar API
extraction/extractor.py	Claude Sonnet NER: batch 10 papers/call, exponential backoff, resumable via .progress.json
extraction/normalizer.py	Entity name → canonical ID: HGNC alias table → PubChem → REST fallback
graph/builder.py	Build NetworkX DiGraph from entities.jsonl; upsert nodes+edges; citation-weighted confidence
graph/query.py	Graph traversal: expand_query_entities(), find_trials_for_target(), get_entity_evidence()
graph/serializer.py	Save/load graph: pickle (fast load at startup) + JSON (human-readable export)
rag/indexer.py	Build ChromaDB collection; section-aware chunking; citation_count in metadata
rag/retriever.py	search(), search_by_entities(), citation-weighted re-ranking
agents/research_agent.py	Multi-step synthesis agent (streaming): entity extraction → KG expansion → RAG → synthesis
app.py	Gradio UI (tabbed: Ask + Therapy Landscape): loads graph + ChromaDB + landscape once at startup
landscape.py	Therapy Landscape rendering: Plotly sunburst + detail-panel HTML from landscape.json
main.py	CLI interface (Rich console)
scripts/	Offline pipeline scripts: run once in order (ingest → extract → build_graph → build_index → build_landscape)

Offline Pipeline Run Order

Run these once to build the knowledge assets. Each is resumable.

# 1. Ingest papers from PubMed + PMC full text + Semantic Scholar citation counts
uv run python scripts/ingest_papers.py

# 2. Ingest ALS clinical trials (can run in parallel with step 1)
uv run python scripts/ingest_trials.py

# 3. Extract entities from papers using Claude (resumable — safe to interrupt)
uv run python scripts/extract_entities.py

# 4. Build knowledge graph
uv run python scripts/build_graph.py

# 5. Build ChromaDB vector index
uv run python scripts/build_index.py

# 6. Build the experimental therapy landscape (offline LLM classification via Batch API)
uv run python scripts/build_landscape.py

# 7. run application
uv run gradio app.py

Key Invariants

• Node key = canonical_id, never raw entity name. Two papers mentioning "TDP-43" and "TARDBP" must produce one node.
• ChromaDB metadata values must be scalars (str/int/float). Lists → comma-separated strings, deserialized on retrieval.
• KG expansion precedes RAG retrieval in the agent loop. Never query ChromaDB with the raw user question alone.
• All heavy compute is offline. No PubMed/extraction calls at query time.

Data File Locations

data/papers/papers.jsonl          — 500 ALS paper records (ALSPaper)
data/trials/trials.jsonl          — ALS clinical trial records
data/extracted/entities.jsonl     — per-paper NER output
data/extracted/canonical_ids.json — entity name → canonical ID registry
data/extracted/.progress.json     — extraction resumability tracker
data/graph/als_graph.pkl          — NetworkX DiGraph (fast load)
data/graph/als_graph.json         — human-readable graph export
data/chroma/                      — ChromaDB SQLite store
data/tools/                       — Claude tool input schemas (JSON)
data/seeds/therapy_classes.json   — ALS mechanism taxonomy for the therapy landscape
data/seeds/therapy_gold.json      — gold-labeled therapies for classifier eval/calibration
data/landscape/landscape.json     — experimental therapy landscape (offline-built, committed to git)

Environment Variables

Variable	Required	Default	Purpose
ANTHROPIC_API_KEY	Yes	—	Claude API
ENTREZ_EMAIL	Yes	—	NCBI Entrez (required by NCBI)
NCBI_API_KEY	No	—	Raises Entrez rate limit 3→10 req/s
LLM_PROVIDER	No	anthropic	Switch to openai
CANDLE_LOG_LEVEL	No	WARNING	Console log verbosity
CANDLE_LOG_DIR	No	logs	Log file directory