SOVEX
CBDC Data Centers Sovereign AI Tokenization Deep Tech Architecture About Team Request access
Sovereign AI / Extensions / Retrieval and knowledge

Retrieval and knowledge.

A model that reasons is not enough; a sovereign model must cite. Grounding binds every answer to the nation's own statutes, registries, and records — with the source attached.

A nation's laws and records become a governed knowledge base

Retrieval starts by turning heterogeneous official sources into a structured, versioned corpus.

01

Authoritative sources

Statutes, regulations, gazettes, court decisions, land and company registries, and supervisory filings are ingested from systems of record — not scraped copies.

02

Structure preserved

Documents are chunked along their real structure — article, clause, schedule — so a retrieved passage carries its citation coordinates, not just a paragraph of text.

03

Temporal versioning

Every provision is stored with its effective dates. The corpus can answer as of a date, distinguishing law in force today from a version since repealed.

04

Ledger-anchored ingestion

Each ingested document's hash is written to the tamper-evident chain, so the corpus can be proven un-altered since a point in time.

Hybrid retrieval finds the right passage, then earns the right to be used

Grounding quality comes from combining lexical precision, semantic recall, and reranking.

01

Lexical and dense

BM25-style exact matching catches statute numbers and defined terms; dense vector search catches paraphrase and concept. The two are fused before ranking.

02

Cross-encoder reranking

A reranker reorders candidates against the actual question, pushing the genuinely responsive clause above merely similar text.

03

In-nation indexes

Embeddings and indexes are computed and stored on sovereign infrastructure. No query or document is embedded through an external API.

04

Query decomposition

Multi-part legal questions are split into sub-queries and retrieved independently, so a compound question does not collapse to a single weak match.

Every claim carries a citation or it is not made

The model is constrained to answer from retrieved sources and to show its work.

01

Citation-bound generation

Answers are assembled from retrieved passages with inline references to instrument and article. A claim without support is withheld rather than invented.

02

Abstention on gaps

When the corpus does not contain an answer, the model says so and points to the nearest authority, instead of filling the void with plausible fiction.

03

Provenance to source

Each citation resolves to the exact registry record or clause, with its effective date and ledger hash, so a reviewer can verify the ground truth directly.

Retrieval respects the classification of what it retrieves

A user only ever sees grounding drawn from records they are cleared to read.

01

Query-time filtering

Access policy is applied during retrieval, not after generation, so restricted records never enter the model's context for an uncleared user.

02

Per-record classification

Chunks inherit the classification and need-to-know of their source. A sealed filing cannot ground an answer for someone without clearance.

03

Audited access

Every retrieval — who asked, what was returned — is logged to the ledger, producing a reviewable trail of which records informed which answer.

The law changes and the knowledge base changes with it

Grounding is only trustworthy if it reflects the current state of official record.

01

Incremental updates

New gazette entries, amendments, and registry changes are ingested continuously. A repeal propagates so the model stops citing dead law.

02

No stale weights

Knowledge lives in the retrievable corpus, not baked into weights, so updating the law is a data operation rather than a retraining project.

03

Change traceability

Every corpus revision is versioned and hash-chained, so an answer given last quarter can be reproduced against the corpus as it stood then.

Build it sovereign.

Talk to us about retrieval and knowledge in a sovereign deployment.