QUESTION + SEARCH
Find possible records
A user asks a question. Your existing SQL, keyword search, vector database, or RAG framework returns candidates. At this point they are possible matches, not established truth.
How it works
Four steps from business question to controlled answer.
PAIL adds one evidence checkpoint. Your files, database, retriever, and chosen model can remain in place.
PAIL BOUNDARY
Read the exact scope
PAIL identifies the requested transaction, order, customer, case, trace, time, and source boundary. Unrelated neighbors are not allowed to redefine the question.
EVIDENCE CHECK
Compare relations and sources
PAIL checks whether the rows agree on fields that matter to the question. A disagreement is kept as evidence instead of being averaged into a ranking score.
DECISION
Release, report, or refuse
Verified facts are packaged with their sources. Conflicts are reported. Missing evidence produces a refusal. Only a verified packet is eligible for AI generation.
Who is responsible for what?
A narrow role makes the system easier to trust.
Store the truth
CRM, payment, ERP, observability, and document systems remain the sources of record.
Find candidates
SQL, FTS, vectors, LangChain, or another search layer locates possible evidence.
Control evidence
Checks identity, relation, conflict, provenance, unsafe fields, and packet completeness.
Explain or review
The model writes from verified facts; a human resolves conflicts and uncertain cause.
Integration
Put the decision between retrieval and generation.
PAIL can read local files itself or receive candidate documents from an existing RAG application.
question
-> your_retriever.search()
-> pail.verify(candidates)
-> VERIFIED_PACKET
? your_llm.generate(packet)
: show_conflict_or_refusal()See it yourself
Run the same records with and without the checkpoint.
The public demo uses synthetic data and exposes no protected runtime rules.