Opportunity Relay White-label Next.js & AI delivery Book a 20-minute fit call

Project 01 / Applied AI

Evidence Desk

A RAG chatbot that knows when not to answer.

Ask in English or Spanish. Every supported response has inspectable evidence; anything the approved sources cannot prove is refused, not improvised.

Open the live lab See the guardrails
  • 01Bilingual requests
  • 02Citation contract
  • 03Fail-closed output
Answer route Evidence before eloquence

The problem

A fluent answer is not the same as a reliable one.

Generic chat interfaces make confidence difficult to inspect. When a model fills a gap with plausible language, the person reading it has to discover the mistake later.

Evidence Desk changes the contract: retrieve first, answer only from those passages, attach the sources, and expose why the system answered or declined.

Input
A plain-language question
Allowed context
Retrieved approved sources
Safe fallback
Explicit abstention

Live RAG lab

Ask. Inspect. Verify.

This interface calls the project endpoint directly. The answer, citations and retrieval trace below come from the returned response—not from decorative sample data.

02 / Conversation Grounded response
Verified context 0 / 8 Ready

Ready for a question.

  1. Evidence DeskReady

    Choose a starter question or write your own. I will show the supporting passages, or clearly say that the available sources are not enough.

Architecture

A short path from question to accountable answer.

Each stage narrows what the next stage is allowed to do. The model is one component in a controlled retrieval system—not the source of record.

  1. 01

    Ingest

    Accept only reviewed source material and keep its document metadata.

  2. 02

    Index

    Split content into retrievable passages with titles, sections and source URLs.

  3. 03

    Retrieve

    Match the question to relevant passages and record what was returned.

  4. 04

    Constrain

    Generate within the retrieved evidence and the requested response language.

  5. 05

    Verify

    Require citations for support; otherwise return an explicit abstention.

Anti-hallucination guardrails

Useful only when the boundary is visible.

The safest response is sometimes no response. These controls make that decision legible to both the person asking and the person maintaining the system.

  1. 01

    Approved corpus only

    Retrieval is limited to the source set selected for the project. General model knowledge is not treated as evidence.

  2. 02

    Evidence threshold

    If the retrieved passages do not support the requested claim, the output must abstain rather than complete the gap.

  3. 03

    Citation contract

    A supported answer must include inspectable citations with a source title, section, URL and excerpt.

  4. 04

    Untrusted content stays untrusted

    Instructions embedded inside retrieved documents are content to quote or ignore, never commands that override system policy.

  5. 05

    Fail-closed interface

    If the API claims support but omits evidence, this page displays a safe refusal instead of presenting the answer as grounded.

  6. 06

    Visible trace

    Retrieved count, top score, abstention reason and matched terms remain available for human review.

Testing protocol

Three perspectives. The same evidence standard.

The interface is designed to be tested for technical failure, procedural accuracy and first-use clarity. Results are inspectable in the response and trace; this page does not invent pass rates.

PerspectiveQuestions usedWhat to inspectSafe outcome
DeveloperMalformed payloads, timeouts, unsafe citation URLs and conflicting flags.Schema validation, abort handling, DOM text rendering and fail-closed state.No unsafe link, stale loading state or unsupported positive answer.
Person with a manualSpecific process and policy questions that can be checked against a known section.Answer wording, cited excerpt, section label and source destination.The claim is directly supported by the displayed passage.
First-time visitorPlain-language presets, an out-of-scope question and a retry after failure.Language choice, status labels, recovery path and abstention clarity.The next action is understandable without hidden instructions.

Build from your sources

Turn a manual, knowledge base or policy set into answers people can verify.

Start with one high-value question set and a clearly owned source of truth.

Discuss a grounded assistant Back to all projects