Sourceform technology

Context, found inside a boundary.

Retrieval starts with what this exact deployment is allowed to know. Only then does Sourceform understand the question, combine retrieval routes, and verify the answer against original evidence.

Permission before relevance

Search cannot grant access.

Search infrastructure is a derived, untrusted index. Every result is replaced with its immutable source evidence and checked against the tenant and publication snapshot before a reranker or model can see it.

  1. 01Authorize deployment
  2. 02Retrieve candidates
  3. 03Reauthorize evidence
  4. 04Allow model egress
Context retrieval

Several signals. One evidence set.

The retrieval route is bounded and deterministic. It may expand the search, but it never expands authorization.

Hybrid retrievalLexical + semantic

Exact words and intended meaning both matter.

BM25Terms, names, identifiers
DenseMeaning and paraphrase
Reciprocal rank fusionCombine rankings without trusting raw score scales
RerankingOrder the authorized evidence for this question
Question frame

Understand the request without widening it.

  • DecompositionUp to four explicit subquestions
  • Entities and aliasesTenant-scoped; ambiguity stays visible
  • Coreference cuesReferences such as “it” or “that policy” are flagged
Relation route

Multi-hop, but never open-ended.

Concepts, entities, aliases, and relations live in an immutable Semantic Release. Only verified relations with original evidence can contribute context, and traversal stops after two hops.

Atlas 2024SUPERSEDED BYAtlas 2026
Relation evidence returns to the standard authorization path.
Time and versions

“Current” is a question, not a pointer.

Years, “latest,” and effective-version language become explicit temporal intent. Answering still reads only the exact Source Versions frozen into the Bot Deployment.

Source Version
Immutable
Publication Snapshot
Exact
Mutable current pointer
Never read
Query routes

Different questions need different context paths.

The route changes. The authorized tenant, bot deployment, publication snapshot, and source versions do not.

QuestionWhat does SKU ATLAS-24 include?
01 · UnderstandQuoted values and identifiers stay exact.
02 · RouteExact hybridBM25 keeps the identifier; dense retrieval finds explanatory passages.
03 · RejoinRRF combines both rankings before reranking.
Evidence preservation

The answer can point back to the original.

Normalization creates searchable structure without replacing the original location. Each eligible passage retains its Source Version and the locator appropriate to its format.

DocumentPage · section · source versionOriginal reading order remains evidence.
WebCanonical URL · heading · source versionThe fetched version—not the future live page—is cited.
TableSheet · range · source versionQueries compile to allowlisted operations; free SQL is rejected.
Claim-level verification

Generation is a draft, not a decision.

Every atomic claim must cite evidence that survived authorization. Literal facts such as dates and numbers are checked against those passages, then the verifier decides what can be delivered.

Review the trust boundary
Claim outcome
SUPPORTED

May be deliveredWith authorized citations

CONTRADICTED

Removed or surfaced as conflictNever silently reconciled

INSUFFICIENT

Removed from the answerIf nothing remains, Sourceform abstains

A partial answer contains only supported claim text. Provider or verifier failure never falls back to an uncited answer.
Honest outcomes

Not every question should become an answer.

Conflict

When verified published sources disagree, Sourceform preserves both citations and asks for resolution instead of choosing.

Partial answer

Supported claims may be delivered while rejected claim text is removed completely.

Abstention

If eligible evidence is insufficient, the answer says so. An abstention is not billed as a verified answer.

A reproducible answer system

Publish knowledge deliberately. Answer from that exact release.

Build your assistant