feat(iace): FMEA source-document register + Anthropic extraction (Haiku)

Quote-verifiable failure extraction via Claude (Haiku 4.5): PDF sent
directly, tool-schema forces verbatim source quotes + applicable flag +
confidence — replaces the unreliable local llama run. Only applicable=true
tuples ingest into bp_iace_failure_kb; every processed doc lands in the
source manifest.

Frontend: FMEA tab now shows a "Quelldokumente" register (every document we
use, with source + licence + link + what was extracted) served from the
embedded manifest via GET /iace/failure-knowledge/sources. Manifest is
placeholder until the 100-doc Haiku run is folded in.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
This commit is contained in:
Benjamin Admin
2026-06-13 13:34:41 +02:00
parent f1ac45dacf
commit 445079cfb2
9 changed files with 395 additions and 0 deletions
@@ -0,0 +1,187 @@
#!/usr/bin/env python3
"""Quote-verifiable FailureKnowledge extraction via the Anthropic API.
Sends each NASA NTRS PDF DIRECTLY to Claude (native PDF + vision, incl. scanned)
and forces a structured tool call whose schema REQUIRES a verbatim source quote
for the key fields — every value is auditable against the document. Only
public-reuse-licensed NTRS docs are processed. applicable=true tuples are
ingested into bp_iace_failure_kb; EVERY processed doc is recorded in the source
manifest (used vs checked-not-used) for the FMEA frontend register.
Env: ANTHROPIC_API_KEY (req), ANTHROPIC_MODEL (default claude-haiku-4-5-20251001),
MAX_DOCS (default 100), RAG (default https://127.0.0.1:8097).
Out: /tmp/fmea_harvest/anthropic.jsonl, /tmp/fmea_harvest/nasa_failure_sources.json
"""
import base64, json, os, ssl, subprocess, sys, time, urllib.parse, urllib.request, urllib.error
API = "https://api.anthropic.com/v1/messages"
NTRS = "https://ntrs.nasa.gov"
RAG = os.environ.get("RAG", "https://127.0.0.1:8097")
COLLECTION = "bp_iace_failure_kb"
MODEL = os.environ.get("ANTHROPIC_MODEL", "claude-haiku-4-5-20251001")
KEY = os.environ.get("ANTHROPIC_API_KEY", "")
MAX_DOCS = int(os.environ.get("MAX_DOCS", "100"))
WORK = "/tmp/fmea_harvest"
OUT = f"{WORK}/anthropic.jsonl"
MANIFEST = f"{WORK}/nasa_failure_sources.json"
QUERIES = ["lessons learned failure", "lessons learned anomaly valve seal bearing motor",
"failure investigation fracture leak short corrosion", "reliability failure mechanism component"]
os.makedirs(WORK, exist_ok=True)
_ctx = ssl.create_default_context()
_ctx.check_hostname = False
_ctx.verify_mode = ssl.CERT_NONE
TOOL = {
"name": "record_failure",
"description": "Record the hardware failure described in the document, with verbatim source quotes.",
"input_schema": {"type": "object", "properties": {
"applicable": {"type": "boolean", "description": "true ONLY if the document describes a concrete hardware/component failure"},
"component": {"type": "string"}, "component_quote": {"type": "string", "description": "verbatim sentence naming the component"},
"failure_mode": {"type": "string"}, "failure_mode_quote": {"type": "string", "description": "verbatim sentence describing how it failed"},
"mechanism": {"type": "string"}, "effect": {"type": "string"},
"hazard": {"type": "string", "enum": ["mechanical_hazard", "electrical_hazard", "thermal_hazard", "fire_explosion", "pneumatic_hydraulic", "none"]},
"control": {"type": "string"}, "confidence": {"type": "string", "enum": ["high", "medium", "low"]}},
"required": ["applicable", "component", "failure_mode", "component_quote", "failure_mode_quote", "confidence"]},
}
PROMPT = ("Read this NASA engineering document and extract the single primary hardware failure. Use the "
"record_failure tool. Quote the SOURCE SENTENCE verbatim for component and failure_mode. If the "
"document does not describe a concrete component failure, call the tool with applicable=false. "
"Never invent values — leave a field empty rather than guess.")
def get_json(url, data=None, headers=None, timeout=120):
req = urllib.request.Request(url, data=data, headers=headers or {})
with urllib.request.urlopen(req, timeout=timeout, context=_ctx) as r:
return json.load(r)
def ntrs_usable(l):
if str(l.get("distribution", "")).upper() != "PUBLIC":
return False
ec = l.get("exportControl") or {}
if any(str(ec.get(k, "")).upper() == "YES" for k in ("isExportControl", "ear", "itar")):
return False
if (l.get("cui") or {}).get("isCui"):
return False
cp = l.get("copyright") or {}
if cp.get("containsThirdPartyMaterial"):
return False
return str(cp.get("determinationType", "")).upper() in ("PUBLIC_USE_PERMITTED", "GOV_PUBLIC_USE_PERMITTED")
def pdf_url(l):
for d in l.get("downloads", []):
if "pdf" in str(d.get("mimetype", "")).lower():
o = (d.get("links") or {}).get("original")
if o:
return NTRS + o
return None
def extract(pdf_b64):
body = json.dumps({"model": MODEL, "max_tokens": 1024, "tools": [TOOL],
"tool_choice": {"type": "tool", "name": "record_failure"},
"messages": [{"role": "user", "content": [
{"type": "document", "source": {"type": "base64", "media_type": "application/pdf", "data": pdf_b64}},
{"type": "text", "text": PROMPT}]}]}).encode()
hdr = {"x-api-key": KEY, "anthropic-version": "2023-06-01", "content-type": "application/json"}
for attempt in (1, 2):
try:
resp = get_json(API, body, hdr, timeout=180)
for b in resp.get("content", []):
if b.get("type") == "tool_use":
return b.get("input"), resp.get("usage", {})
return None, resp.get("usage", {})
except (urllib.error.URLError, ConnectionError) as e:
if attempt == 2:
raise
time.sleep(3)
return None, {}
def ingest(did, title, t, lic, url):
md = (f"# NASA Lesson {did}: {title}\n\n- Source: NASA NTRS {did}\n- License: {lic}\n- URL: {url}\n"
f"- verified: false (Claude-extracted, quote-checked)\n\n"
f"Component: {t.get('component','')}\nFailure mode: {t.get('failure_mode','')}\n"
f"Mechanism: {t.get('mechanism','')}\nEffect: {t.get('effect','')}\nHazard: {t.get('hazard','')}\n"
f"Control: {t.get('control','')}\nConfidence: {t.get('confidence','')}\n\n"
f"Component quote: {t.get('component_quote','')}\nFailure-mode quote: {t.get('failure_mode_quote','')}\n")
p = f"{WORK}/fk_{did}.md"
open(p, "w").write(md)
meta = json.dumps({"title": f"NASA NTRS {did}: {title}"[:120], "license": lic, "source": f"NASA NTRS {did}", "verified": "false"})
r = subprocess.run(["curl", "-sk", "--max-time", "90", "-X", "POST", f"{RAG}/api/v1/documents/upload",
"-F", f"file=@{p}", "-F", f"collection={COLLECTION}", "-F", "data_type=failure_kb",
"-F", "use_case=iace_fmea", "-F", "year=2024", "-F", f"metadata_json={meta}"],
capture_output=True, text=True, timeout=120)
try:
os.remove(p)
except Exception:
pass
return "chunks_count" in r.stdout
def main():
if not KEY:
print("ERROR: ANTHROPIC_API_KEY not set", file=sys.stderr); sys.exit(1)
subprocess.run(["curl", "-sk", "-X", "POST", f"{RAG}/api/v1/collections", "-H", "Content-Type: application/json",
"-d", json.dumps({"name": COLLECTION, "vector_size": 1024})], capture_output=True)
manifest, seen, processed = [], set(), 0
for q in QUERIES:
if processed >= MAX_DOCS:
break
for frm in (0, 100, 200):
if processed >= MAX_DOCS:
break
try:
res = get_json(f"{NTRS}/api/citations/search?q={urllib.parse.quote(q)}&page.size=100&page.from={frm}&highlight=false", timeout=90)
except Exception as e:
print(f"search '{q}'@{frm} error: {e}", flush=True); continue
for l in res.get("results", []):
if processed >= MAX_DOCS:
break
did = l.get("id")
if not did or did in seen:
continue
if not ntrs_usable(l):
continue
url = pdf_url(l)
if not url:
continue
seen.add(did)
pdf = f"{WORK}/{did}.pdf"
subprocess.run(["curl", "-sL", "--max-time", "150", "-o", pdf, url], capture_output=True)
sz = os.path.getsize(pdf) if os.path.exists(pdf) else 0
if sz < 1000 or sz > 30_000_000:
try:
os.remove(pdf)
except Exception:
pass
continue
b64 = base64.b64encode(open(pdf, "rb").read()).decode()
os.remove(pdf)
try:
t, usage = extract(b64)
except Exception as e:
print(f" {did} extract error: {e}", flush=True); continue
if not t:
continue
title = (l.get("title") or "")[:110]
lic = f"Public Domain (NASA NTRS, {(l.get('copyright') or {}).get('determinationType','')})"
used = bool(t.get("applicable"))
t.update({"_id": did, "_title": title, "_license": lic, "_url": url, "_model": MODEL, "_used": used})
open(OUT, "a").write(json.dumps(t, ensure_ascii=False) + "\n")
if used:
ingest(did, title, t, lic, url)
manifest.append({"id": did, "title": title, "source": f"NASA NTRS {did}", "license": lic, "url": url,
"used": used, "component": t.get("component", ""), "failure_mode": t.get("failure_mode", ""),
"confidence": t.get("confidence", "")})
processed += 1
json.dump({"generated": "nightly", "model": MODEL, "count": len(manifest), "documents": manifest},
open(MANIFEST, "w"), ensure_ascii=False, indent=1)
print(f" [{processed}] {did} used={used} {t.get('component','?')}{t.get('failure_mode','?')} ({usage.get('input_tokens')}in)", flush=True)
used_n = sum(1 for m in manifest if m["used"])
print(f"DONE: {processed} processed, {used_n} used (applicable) -> {MANIFEST}", flush=True)
if __name__ == "__main__":
main()