Structura.
The evidence layer for production AI. Chunks are derived artifacts. Atoms are the truth. Structura keeps source truth alive as addressable atoms, verifies every claim against its evidence, and only lets an answer, agent, or tool act after the context is proven.
Source truth to proven answer.
Ingestion produces atoms with locators — page + bounding box, row + cell, JSON path. The canonical evidence store is the source of truth; every index is a rebuildable projection. Watch the pipeline run:
Tamper-evident, by design.
Every decision carries its evidence chain and hashes. The ledger gives inclusion and consistency proofs — honest engineering, not blockchain magic.
- BLAKE3Canonical content hash — config_hash + run_hash on every artifact
- Ed25519Signature sidecar for the .structura envelope
- Merkle ledgerRFC-6962-style append-only log with inclusion and consistency proofs
- SimHash · MinHashNear-duplicate detection; perceptual aHash catches same-scan-different-file
- PQC migration planEd25519 -> ML-DSA / SLH-DSA hybrid path, sketched in structura.trust
Benchmarks with provenance.
From the repo's own bench reports — every number ships with the config hash that produced it.
+0.244 F1 over the fuzzy-matching baseline — with full provenance on every decision.
Full 50-document corpus parsed in 0.52 s total — Rust engine, deterministic output.
What the fabric guarantees.
Atom-native retrieval
Page boxes, table cells, figures, and formulas are first-class evidence — not loose chunks. Chunks are derived artifacts, never the source of truth.
Hybrid retrieval with RRF fusion
Lexical BM25 + dense vectors + learned sparse + multivector late interaction, fused with reciprocal rank fusion and cross-encoder reranking.
Claim-level verification
Every cited atom must actually support each generated claim. If the evidence doesn't prove it, the system abstains instead of guessing.
Trust-tiered provenance
Signed internal evidence outranks uploads, approved web, scraped web, and generated memory. Trust is explicit, not implied.
Deterministic, replayable runs
Every artifact carries a BLAKE3 config_hash and run_hash — the same inputs and config reproduce the same decision, bit for bit.
Governed semantic cache
Cache hits are candidates, not answers — they must re-pass freshness, permission, evidence-overlap, and claim-support checks.
Freshness & invalidation
Ingest time, valid time, and source version gate stale context before it ever reaches a prompt.
GraphRAG multi-hop
Graph paths across entities and records, with every hop tied back to atomic evidence.
Evidence packets
Portable governed context — locators, redaction state, policy refs, prompt and context hashes — ready for any RAG, LLM, or agent runtime.
“Structura doesn't just retrieve more context. It retrieves atomic evidence, verifies each claim, tracks freshness and trust — and only then lets an answer proceed.”