refactor(prompts): consolidate system prompts + docs/lockfile housekeeping (#23)

* refactor(prompts): consolidate all system prompts into prompts.py

extraction/extractor.py and ingestion/clinicaltrials.py each defined their
own local system prompt, and prompts.py held a stale, unused EXTRACTION_SYSTEM
that had drifted from the one the extractor actually ran. Make prompts.py the
single source of truth: move the live extraction and trial-target prompts there
(byte-identical to what ran, so extraction output is unchanged) and import them.

Now all four system prompts (EXTRACTION_SYSTEM, TRIAL_EXTRACTION_SYSTEM,
LANDSCAPE_SYSTEM, SYNTHESIS_SYSTEM) live in prompts.py, matching CLAUDE.md.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* chore: doc paper count 500→10,000 and sync uv.lock to pinned gradio

CLAUDE.md described the corpus as ~500 papers; it is now ~10,000.
uv.lock had drifted to gradio 6.20.0 while pyproject pins gradio==6.14.0 —
re-lock so the lockfile matches the committed pin.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

CLAUDE.md

@@ -2,7 +2,7 @@
 
 ## 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.
+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 ~10,000 curated ALS papers, enriched by a knowledge graph.
 
 **Sibling project**: beacon (patient-facing clinical trial finder at `../beacon`). Follow the same conventions.
 
@@ -10,7 +10,7 @@ Candle-fire is a physician-facing ALS research intelligence tool. A physician as
 
 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.
+2. **RAG (Vector Search)**: ChromaDB collection of ~10,000 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.
 

extraction/extractor.py

@@ -30,6 +30,7 @@ from config import (
 from extraction.normalizer import CanonicalRegistry, guess_entity_type, normalize_entity
 from logging_config import get_logger
 from models import ALSPaper, ExtractedEntity, EntityRelationship, PaperExtractionResult
+from prompts import EXTRACTION_SYSTEM
 from tools import EXTRACTION_TOOLS
 
 _logger = get_logger("extraction.extractor")
@@ -38,17 +39,6 @@ _logger = get_logger("extraction.extractor")
 # hour; the ceiling is 24h.
 _POLL_INTERVAL_S = 30
 
-_EXTRACTION_SYSTEM = """\
-You are a biomedical NLP expert specializing in ALS (amyotrophic lateral sclerosis).
-Extract entities and relationships from each paper using the extract_entities tool.
-Call it once per paper. Use the full text when provided — it is richer than the abstract alone.
-
-Entity types: Gene, Protein, Compound, Pathway, Phenotype, Mechanism.
-Relationship types: BINDS, INHIBITS, ASSOCIATED_WITH, TESTED_IN, EXPRESSED_IN, CO_OCCURS.
-
-Be precise. Only extract entities explicitly mentioned. Return pmid exactly as given.
-"""
-
 
 def extract_all(
     papers_path: Path = PAPERS_PATH,
@@ -224,7 +214,7 @@ def _build_params(batch: list[ALSPaper]) -> MessageCreateParamsNonStreaming:
     return MessageCreateParamsNonStreaming(
         model=EXTRACTION_MODEL,
         max_tokens=8192,
-        system=_EXTRACTION_SYSTEM,
+        system=EXTRACTION_SYSTEM,
         tools=list(EXTRACTION_TOOLS),
         tool_choice={"type": "any"},
         messages=[{"role": "user", "content": _format_batch(batch)}],

ingestion/clinicaltrials.py

@@ -8,6 +8,7 @@ import httpx
 
 from config import CTGOV_BASE, EXTRACTION_MODEL
 from logging_config import get_logger
+from prompts import TRIAL_EXTRACTION_SYSTEM
 
 if TYPE_CHECKING:
     import anthropic
@@ -16,16 +17,6 @@ _logger = get_logger("ingestion.clinicaltrials")
 
 _TRIAL_BATCH_SIZE = 10
 
-_TRIAL_EXTRACTION_SYSTEM = """You are a biomedical NLP expert specializing in ALS (amyotrophic lateral sclerosis) clinical trials.
-
-For each trial provided, identify the primary biological target(s) being tested or modulated:
-- Genes silenced or corrected (e.g., SOD1, TARDBP, FUS, C9orf72, NEK1, VCP, TBK1)
-- Proteins targeted (use canonical gene symbol, e.g. TARDBP for TDP-43 protein)
-- Compounds/drugs — report the molecular or pathway target, not the drug name (e.g., a trial of AMX0114 targets TARDBP)
-- Mechanisms (e.g., neuroinflammation, oxidative stress, glutamate excitotoxicity)
-
-Call extract_trial_targets once per trial. Return an empty targets list only when no specific molecular or mechanistic target is identifiable."""
-
 
 def fetch_als_trials(
     client: "anthropic.Anthropic | None" = None,
@@ -204,7 +195,7 @@ def _call_claude_batch(
             response = client.messages.create(
                 model=EXTRACTION_MODEL,
                 max_tokens=4096,
-                system=_TRIAL_EXTRACTION_SYSTEM,
+                system=TRIAL_EXTRACTION_SYSTEM,
                 tools=tools,
                 tool_choice={"type": "any"},
                 messages=[{"role": "user", "content": "\n".join(lines)}],

prompts.py

@@ -1,29 +1,34 @@
-"""System prompts for extraction and synthesis agents."""
+"""System prompts for every Claude call in candle-fire.
 
+Single source of truth: entity extraction (extraction/extractor.py), trial-target
+extraction (ingestion/clinicaltrials.py), therapy-landscape classification
+(scripts/build_landscape.py), and query synthesis (agents/research_agent.py) all import
+their system prompt from here. Add or modify prompts in this file only.
+"""
+
+# Paper entity/relationship extraction — extraction/extractor.py (Batch API, one call/paper).
 EXTRACTION_SYSTEM = """\
-You are a biomedical NLP expert specializing in ALS (amyotrophic lateral sclerosis) research.
-Your task is to extract biomedical entities and relationships from ALS paper abstracts.
-
-Entity types to extract:
-- Gene: genetic loci (e.g., SOD1, TARDBP, FUS, C9orf72)
-- Protein: protein products (e.g., TDP-43, FUS protein, SOD1 protein)
-- Compound: drugs, small molecules, biologics (e.g., riluzole, tofersen, AMX0035)
-- Pathway: biological pathways or processes (e.g., glutamate excitotoxicity, autophagy)
-- Phenotype: disease features or clinical observations (e.g., bulbar onset, respiratory failure)
-- Mechanism: molecular or cellular mechanisms (e.g., protein aggregation, oxidative stress)
-
-Relationship types to extract:
-- BINDS: compound/protein binds to a target
-- INHIBITS: compound/gene inhibits a target
-- ASSOCIATED_WITH: entity is associated with a disease phenotype or another entity
-- TESTED_IN: compound is tested in a clinical trial or animal model
-- EXPRESSED_IN: gene/protein is expressed in a tissue or cell type
-- CO_OCCURS: entities frequently co-occur in ALS context (weakest relationship)
-
-Be precise. Only extract entities explicitly mentioned. Confidence reflects how clearly
-the entity is identified in the text (1.0 = unambiguous, 0.5 = inferred, 0.3 = uncertain).
+You are a biomedical NLP expert specializing in ALS (amyotrophic lateral sclerosis).
+Extract entities and relationships from each paper using the extract_entities tool.
+Call it once per paper. Use the full text when provided — it is richer than the abstract alone.
+
+Entity types: Gene, Protein, Compound, Pathway, Phenotype, Mechanism.
+Relationship types: BINDS, INHIBITS, ASSOCIATED_WITH, TESTED_IN, EXPRESSED_IN, CO_OCCURS.
+
+Be precise. Only extract entities explicitly mentioned. Return pmid exactly as given.
 """
 
+# Clinical-trial target extraction — ingestion/clinicaltrials.py (one call/trial).
+TRIAL_EXTRACTION_SYSTEM = """You are a biomedical NLP expert specializing in ALS (amyotrophic lateral sclerosis) clinical trials.
+
+For each trial provided, identify the primary biological target(s) being tested or modulated:
+- Genes silenced or corrected (e.g., SOD1, TARDBP, FUS, C9orf72, NEK1, VCP, TBK1)
+- Proteins targeted (use canonical gene symbol, e.g. TARDBP for TDP-43 protein)
+- Compounds/drugs — report the molecular or pathway target, not the drug name (e.g., a trial of AMX0114 targets TARDBP)
+- Mechanisms (e.g., neuroinflammation, oxidative stress, glutamate excitotoxicity)
+
+Call extract_trial_targets once per trial. Return an empty targets list only when no specific molecular or mechanistic target is identifiable."""
+
 LANDSCAPE_SYSTEM = """\
 You are an ALS-pharmacology expert classifying experimental therapies by mechanism of action.
 For each therapy you are given EVIDENCE (its trial summaries + retrieved paper abstracts).

uv.lock

@@ -472,7 +472,7 @@ requires-dist = [
     { name = "anthropic", specifier = ">=0.50.0" },
     { name = "biopython", specifier = ">=1.84" },
     { name = "chromadb", specifier = ">=0.5.0" },
-    { name = "gradio", specifier = ">=6.14.0,<7.0.0" },
+    { name = "gradio", specifier = "==6.14.0" },
     { name = "httpx", specifier = ">=0.27.0" },
     { name = "networkx", specifier = ">=3.3" },
     { name = "openai", specifier = ">=2.44.0" },
@@ -980,7 +980,7 @@ wheels = [
 
 [[package]]
 name = "gradio"
-version = "6.20.0"
+version = "6.14.0"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
     { name = "anyio" },
@@ -1013,9 +1013,9 @@ dependencies = [
     { name = "typing-extensions" },
     { name = "uvicorn" },
 ]
-sdist = { url = "https://files.pythonhosted.org/packages/f8/96/3618f8a189180ab3d4e5f8b258fd92e419eda9f49b6ab4dd50c6782cda33/gradio-6.20.0.tar.gz", hash = "sha256:8672866e9225a1f8297325a6742bb525a3a0ddc8aa4d75d99beb10a27ea2915f", size = 44844340, upload-time = "2026-07-07T18:53:57.402Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/de/bd/7d1544571de4566138e50c868b91bb79e38c998266896d38fed3d3a77898/gradio-6.14.0.tar.gz", hash = "sha256:4972ef7d01ac57472772624eb4e095767b6c8f3cd4846b7fea648e8034cda9f8", size = 36026409, upload-time = "2026-04-30T16:50:31.698Z" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/d7/f4/7dc4540836e326391a9d1ba20193d795d5a4582839b60f5c10397f5120bb/gradio-6.20.0-py3-none-any.whl", hash = "sha256:91d2ec29c37917d55b18dba87aa15bb15c4796d84dce9a9ef4ec7d87453a7fee", size = 30964820, upload-time = "2026-07-07T18:53:52.403Z" },
+    { url = "https://files.pythonhosted.org/packages/7e/e2/41f991b2e212b7afda3a9927676ebf8af302e5a2c632b330bd70bf2cf2c1/gradio-6.14.0-py3-none-any.whl", hash = "sha256:bb702f5ab643510d167bae54269ad6e985c2185174d388fe542cc5957f51f4fd", size = 19687959, upload-time = "2026-04-30T16:50:26.914Z" },
 ]
 
 [[package]]