Build a domain RAG (legal, medical)

CiteNexus is built for domains where a wrong answer is worse than no answer — legal, medical, finance, compliance. The recipe is the same across them: cite or abstain, and prove it on a golden set that rewards abstention.

Python facade — Go and JavaScript have no equivalent

CiteNexus and the ingest / ask /evaluate / reconcile / delete verbs, plus the config layer they read, are the Python facade. The Go and JavaScript packages ship the deterministic core instead — the faithfulness gate, tokenizer, chunker, RRF and Result, at conformance parity — and enter through Ask / ask over an in-memory corpus (quickstart). The Python samples on this page have no port tab because there is nothing to put in it.

1. Curate authoritative sources. Ingest the documents you’d actually cite — a statute, a clinical guideline, a policy PDF — not the open web. Idempotent ingest (bulk) means you can re-run as the corpus grows.

   for path in guideline_pdfs:
       rag.ingest(str(path), document_id=path.stem)

2. Keep strict mode. It’s the default, and it’s the guarantee — the answer is generated only from retrieved passages and every claim is verified against its citation (what that check actually proves). For a regulated surface, do not drop to normal/exploratory.

   response = rag.ask("What is the maximum pediatric dose of drug X?")  # strict
   if response.evidence.decision.value == "answered":
       src = response.sources[0]
       print(response.answer, "—", src.document, "p.", src.page)   # cite the exact passage
   else:
       print("abstained:", response.answer)   # no dosage invented

3. Inject your models. Any OpenAI-compatible embedding + LLM (+ reranker, + vision for figures). Keep it local (Ollama) for data-residency, or use a hosted endpoint — see custom endpoints & auth.

4. Prove it with a golden set — and assert the refusals separately. evaluate() scores the answerable rows: put the text you expect quoted in the expected column and read groundedness_rate, citation_rate and expected_support_rate.

   question,expected
   "What notice must a CA landlord give a >1yr month-to-month tenant?","60"

   Must-refuse rows do not belong in that CSV. A blank expected is scored as supported only when the row was answered, so a correct refusal lowers expected_support_rate rather than raising it. Assert abstention directly instead:

   from citenexus.answer.result import Decision
   
   must_refuse = ["What is the maximum security deposit under this statute?"]
   for q in must_refuse:
       assert rag.ask(q).evidence.decision is Decision.refused, q

The order those guards run in matters for a regulated surface — grounding first, then the authority floor, then the per-claim gate, and no stage can promote what an earlier one rejected. How it works lays the pipeline out stage by stage.

A real, measured example: law

The law worked example is exactly this recipe on California landlord–tenant law, run live against Jina + Gemini: 100% groundedness and citation, zero fabricated answers across an 11-question golden set — versus the 17–33% hallucination rate reported for generic legal RAG (Stanford RegLab, 2024, cited as context). It also shows the honest failure mode to watch for.

That is a single live run against a generator that is not reproducible at temperature 0 — an identical re-run moved the rate metrics by a question. The safety metrics (100% groundedness, 100% citation, 0 out-of-jurisdiction citations) held across both. Treat the rates as illustration and the safety numbers as the result.

Turn the authority floor on — retrieval alone ranks by relevance

Retrieval and fusion rank by relevance only, so in a corpus mixing a controlling statute with a blog or another jurisdiction, a low-authority source that repeats the query vocabulary can win the citation. Authority weighting ships (ADR-0004) but is unranked by default: attach tiers at ingest with rag.ingest(..., authority={"authority_tier": "controlling-statute"}) and set a floor, and strict mode withholds everything below it — after grounding, before generation. On the law corpus that took out-of-jurisdiction citations from 4 to 0. Tiers are curator assertions; the library never derives them from prose, and a mis-declared tier is a real failure mode. See Authority — grounding is not standing.

Right authority, wrong subject — still open

The floor orders sources; it cannot catch a passage from the highest-authority source that governs a different subject. A statute’s applicability clause and its operative rule are often separate Evidence Units, and every downstream guard is chunk-local — measured at 73% severance on the law corpus. See Subject scope.

Medical, specifically

The same shape applies: ingest the guideline / formulary / label, keep strict mode so a dose or contraindication is only ever quoted from the source (never synthesized), cite the exact page, and keep a separate list of “not in the guideline” questions asserted against Decision.refused — abstention has to be measured outside evaluate(). In any supported script, answers come back in the clinician’s language while the citation stays verbatim in its own language, so an English answer can carry a non-English quote — see languages.