Provable
Drop a million documents.
Prove every answer.
Cited answers for questions. Ranked chunks for agents. Both signed, both verifiable, both on your machine.
from provable import Pipeline
p = Pipeline.from_documents("./pdfs/")
ans = p.query("data retention policy?")
ans.answer # cited text
ans.citations # source spans
ans.proof # SHA-256 MerkleFrom folder to signed answer in four stages.
Every stage is deterministic, inspectable, and reproducible. No black-box embedding API. No vendor upload. No surprises.
The whole library is one import away.
No vendor upload. No external embedding API. Pure Python, runs on your machine.
from provable import Pipeline
pipeline = Pipeline.from_documents("./my_corpus/")
result = pipeline.query("What does our compliance policy say about data retention?")
# result.answer → cited natural-language answer
# result.citations → exact source spans with doc IDs
# result.proof → SHA-256 Merkle commitment (verifiable in <40ms)
# result.verdict → "ANSWERED" or "ABSTAINED"# Same query against a running Provable server
curl "https://your-host/api/query?q=What+does+our+compliance+policy+say"
# Independently verify any returned proof:
curl -X POST "https://your-host/api/verify" \
-H "content-type: application/json" \
-d '{"query": "...", "retrieved_doc_ids": [...], "proof_signature": "PROOF-..."}'
# → { "valid": true, "verify_ms": 1.8, "reason": "all hashes match" }{
"verdict": "ANSWERED",
"answer": "Data retention policy requires 7-year storage...",
"primary_doc_id": "policies/retention-2024.pdf#p3",
"citations": [
{ "doc_id": "policies/retention-2024.pdf#p3", "score": 9.41 },
{ "doc_id": "compliance/gdpr-summary.md", "score": 7.22 }
],
"proof_signature": "PROOF-A8B3C1D2E4F5G6H7I8J9K0L1",
"latency_ms": 113.4
}Cited answer, or honest abstention.
There is no "best guess." There is no fabricated citation. The set is partitioned.
Your agent shouldn't read 5,000 documents.
Provable picks the 47 that matter.
Modern agents waste their context window on irrelevant text. Provable filters your corpus down to the chunks that actually matter for the decision. Cited, signed, ranked. Drop them straight into the system prompt. The agent decides on verified ground truth.
retention-policy-2024.pdf#p3.
Redact PII after year 5 per gdpr-summary.pdf#p12.
PROOF-A8B3C1D2E4F5G6H7
Point it at your industry's documents.
Dosing answers cited to the FDA monograph.
Clinical AI that cites. Audit trail verifiable in milliseconds.
Pin-cited precedent. No Mata v. Avianca sanctions.
Real cases, real paragraphs. Opposing counsel verifies the citation from the signed proof in under 40 ms.
Every customer-facing answer becomes one cryptographic check.
Cited to the actual disclosure document. Supervision review becomes minutes, not hours.
Every eligibility decision defensible at appeal.
Cited to the actual regulation. The administrative record writes itself.
Your corpus never leaves your host.
Try it on a sample corpus.
Pre-loaded with 518 documents across medical, legal, finance, science, and software. Drop your own folder anytime. Same engine.
Three steps. No magic.
Point it at your corpus.
One folder, any format. Cited answers, signed proofs. Honest abstentions.