SourcePath Labs | Evidence-first AI for structured authority documents

SourcePath helps teams work with regulations and specifications without trusting similarity alone.

SourcePath is designed for document sets where hierarchy, authority, applicability, revision history, and cross-reference structure carry real meaning. Instead of flattening those signals into generic chunks, it preserves them during ingestion and uses them to keep retrieval narrow, traceable, and easier to validate.

Why this approach is different

Authority beats proximity.

The goal is to retrieve what governs, not simply what sounds related.

Ingestion does the hard thinking.

High-value links are resolved up front so the system is less likely to drift at question time.

Explainability is part of correctness.

Evidence, references, and missing targets stay visible so a human can validate quickly.

Where SourcePath fits

SourcePath is built for structured authority domains where semantic similarity alone tends to fail: layered regulations, agency manuals, codes, standards, specifications, and versioned requirements.

Government and HHS

Federal, state, county, and agency materials where applicability, supersession, and operational impact matter.

Engineering and technical specifications

Standards, codes, revisions, and flow-down requirements where conflicting or inherited requirements drive cost and risk.

Change and impact analysis

Questions like what changed, what applies now, and what downstream programs or components are affected.

Current example corpus

16003rows in the Federal SNAP XML example

17700rows in the Federal Medicaid XML example

1467references turned into explicit backlog work

8rows in the orientation example

Federal CFR XML typed hierarchy internal cross-references external placeholders ingestion backlog

Start with the examples

Example of document ingestion

A compact walkthrough of hierarchy extraction, cross-reference handling, and backlog generation.

SNAP guided example

Part-level views from a real Federal volume, with hierarchy and reference overlays rendered as zoomable SVGs.

Medicaid guided example

A second Federal domain with overlapping language pressure but different governing structure.

What to ingest next

The backlog shows downstream target documents discovered by parsing references out of the ingested material.

Product doctrine

Similarity is a hint, not a fact. Explicit beats inferred. Authority beats proximity. Explainability is part of correctness.

That doctrine is what drives the ingestion-first design: spend compute early, preserve structure, demote weak links, and keep the smallest correct evidence set visible to the user.