MLCLabs · Practical tool · English

Find the evidence behind a knowledge-assistant answer

A useful assistant needs an answer boundary: which source supports the response, whether the requesting user may retrieve it, and what to do when evidence is missing.

Free to explore. Worksheet inputs stay in this browser tab; no account or paid API is required.

Conceptual title illustration for MLCLabs
Conceptual title image. Any real observation data is identified separately below.
A working retrieval-and-citation demonstration.

This page searches and returns exact extracts from a small public fixture corpus. It does not call an LLM or present generated prose as a verified answer. Every sample document is public. The role selector demonstrates filtering and is not authentication or document security.

Ask the sample documents

Use the example questions or ask your own. Matching is deterministic term-based retrieval with an evidence-coverage threshold; it has no semantic understanding. Questions outside this small corpus should receive no matching evidence.

The fictional source documents are loading.

Run the evaluation, inspect the failures

The six fixed checks cover source selection, recovery evidence, retrieval permissions, a simulated restricted source and an unanswerable question. Run them in your browser and download the actual results. This small regression suite does not measure production quality, resistance to prompt injection or general language-model accuracy.

No evaluation run yet.

Inspect the complete public corpus

These Project Alder documents were written for this demonstration. They are not client records, real personnel assignments or MGCA operating policies.

Project Alder — release evidence guide — Evidence before launch

Fixture role: public · Source ID: release-guide/launch-evidence

A release review requires a named service owner, an evaluation set tied to the task, permission tests and a recovery rehearsal. Open exceptions must have an owner and an agreed next step.

Project Alder — release evidence guide — Recovery and rollback

Fixture role: public · Source ID: release-guide/recovery

The recovery rehearsal must inspect business state as well as application health. Record what can be retried, what must be reconciled and who can authorise a restart.

Project Alder — access guide — Permissions and retrieval

Fixture role: public · Source ID: access-guide/permissions

The assistant may retrieve only documents available to the requesting user. A citation identifies the document and section used; it does not grant access to other material.

Project Alder — access guide — When evidence is missing

Fixture role: public · Source ID: access-guide/unknown

If the approved documents do not support an answer, the assistant should say that the evidence is missing and route the question to its owner. It must not invent a policy or a result.

Project Alder — team planning note — Illustrative budget owner

Fixture role: team · Source ID: planning-note/budget

In this fictional fixture, the budget owner is Sam Patel. Funding changes require the project sponsor's approval. This example is available only when the demonstration role is Project team.

Download corpus and fixed evaluation cases

What changes in a real system?

Production retrieval needs server-enforced permissions before content reaches the user, owned source updates, task-specific evaluation and an answer-generation policy. A browser role dropdown cannot protect private documents: all fixtures here are intentionally public. Keep absence of evidence explicit, and make the source section easy to inspect.

Read the Microsoft retrieval-augmented generation overview and the NIST Generative AI Profile for the wider system context.

Take the next step with context

Keep the question, assumptions and evidence together. These tools help organise a review; an exported record does not approve a design or certify a system.