Read the source
Extract text and structure from PDFs, documents, logs, tables, databases, APIs, audio transcripts, or other supported formats.
How it works
From raw data to evidence authority.
What a normal RAG pipeline does.
A RAG system stores searchable pieces of source material and gives selected pieces to a generator.
Create useful retrieval units
Split by policy section, paragraph, function, transaction, incident, entity, scene, or another boundary that preserves meaning.
Store searchable representations
Each chunk may be embedded into a vector and stored with its original text and metadata. Exact, keyword, SQL, and graph indexes may coexist.
Find candidates and build context
The question is searched, the top candidates are placed in a context packet, and an LLM usually writes the answer.
The PAIL insertion point
Candidate evidence is not yet trusted evidence.
PAIL evaluates candidates after retrieval and before generation. It may also normalize evidence before indexing and verify claims after generation.
Canonical events
Group interleaved logs by transaction, trace, entity, and time rather than arbitrary line windows.
Evidence guard
Remove wrong entities, expose relation conflicts, and preserve the strongest valid bridge.
Permission gate
Only verified packets may reach an optional model; missing or conflicting evidence closes the gate.
Claim check
Model prose can be rejected when it introduces values or conclusions absent from the packet.
Chunking without confusion
Teams define a policy once. They do not cut every file by hand.
One universal fixed-size chunker is convenient but can destroy the exact boundary the answer depends on.
| Source | Useful boundary | Why |
|---|---|---|
| Policy PDF | Heading and policy section | Keep eligibility, exceptions, and version metadata together |
| Bank/POS logs | Transaction, STAN, RRN, trace, or session | Avoid mixing events from concurrent payments |
| Application incident | Incident plus bounded time window | Preserve sequence without absorbing unrelated traffic |
| SQL/CSV | Entity or record | Keep field relationships and keys intact |
| Code | Function or class | Preserve executable scope |
| Story/prose | Scene or overlapping narrative window | Maintain characters and before/after context |
Concrete decision example
“Can I take 30 annual leaves?” is not the same as “How can I take 30 days off?”
The number is identical. The intent and permitted answer are different.
Request
Employee requests 30 days away.
Available
24 annual-leave days remain.
Shortfall
6 days; unpaid leave allowed with approval.
Partially allowed
24 annual + 6 unpaid, manager approval required.
Explain
Template or LLM restates only that decision.
Three valid ways to use a model.
The trustworthy core remains deterministic in every mode.
Rules decide; a template answers
Best for bounded API responses, policy calculations, offline systems, and workflows where language variation is limited.
Rules decide; the model explains
The model sees the decision object and evidence IDs. It cannot change numbers, status, or permitted alternatives.
The model normalizes language; PAIL keeps authority
A model may map flexible language into a canonical intent, but PAIL verifies that structure against evidence and policy before any action or answer.
See the difference
Use the same records with and without the evidence gate.
The public lab makes the outcome visible without exposing the proprietary runtime.