Local-first knowledge assistant lab

Inspect the knowledge,
not just the answer.

DialogueBase builds a question-answering assistant from your own source material - and keeps the machinery visible. Every answer shows its sources. Every correction becomes a test case. Every unanswered question becomes a gap you can close.

6connected lab views
0answers without visible sources
0visitor emails sent automatically

How it works

An assistant you can audit

Load the knowledge

Add your source material in the Sources view. The knowledge base is inspectable line by line - it is the product, not a hidden index.

Ask, inspect, correct

The playground answers from your sources and shows which ones it used. Wrong answer? Correct it - the correction becomes part of the knowledge.

Prove it with test cases

Conversations convert into regression tests. Change a source, rerun the cases, and see what broke before a visitor does.

Close the gaps

Unanswered visitor questions cluster by topic. Each cluster is a source you could write - the roadmap writes itself.

Inside the agent loop

Ask, cite, correct

One question through the loop the way the playground actually runs it: scored retrieval with citations, flagged-source warnings, and a human correction that outranks retrieval next time.

PLAYGROUNDGROUNDED
Do you support SSO?answered from 2 sources - onboarding.md, security.md
SCORED
[1] security.md - chunk 3SCORE 41
[2] onboarding.md - chunk 5SCORE 22
RETRIEVAL IS SCORED, NOT VIBED
EVERY CLAIM CARRIES A CITATION
PLAYGROUNDAnswers only from your sourcesNo source, no answer - the gap list records it instead
CITATIONS - CHUNK LEVEL1 FLAGGED
[1] pricing.md - chunk 2SCORE 38
[2] onboarding.md - chunk 5SCORE 22
legacy-faq.txt - chunk 1outdated pricing table - review before it answers again
SOURCE WARNING
CHUNK-LEVEL RECEIPTS
FLAGGED SOURCES WARN FIRST
SOURCESFlagged chunks warn before they answerInclude or exclude them from retrieval with one toggle
REVIEW - CORRECTIONLEARNED
"What is the refund window?"unresolved in chat - taught by a human: 30 days, receipt required
CORRECTION
rule: saved corrections answer matching questions before source retrieval runs
HUMANS TEACH THE ASSISTANT
CORRECTIONS OUTRANK RETRIEVAL
CORRECTION SAVEDNext match answers from thisThe playground and the public chat both learn it

Sources

The knowledge is inspectable

Visible

Line by line

Every source the assistant can draw on is listed and readable. No hidden index, no mystery context.

Attribution

Answers cite

Each answer names the sources behind it, so a wrong answer traces back to the exact material.

Local

Stays in the lab

Sources live in browser storage. Export them any time; delete them completely.

Playground

Ask, inspect, correct

Ask questions against your knowledge base. Inspect which sources shaped each answer. When the answer is wrong, fix it on the spot - corrections feed back into the knowledge the next answer uses.

Review

Conversations become test cases

Promote real conversations into a regression suite. After any source edit, rerun the cases and compare. Quality becomes something you verify, not something you feel.

Analytics

The gaps behind the score

Clusters

Unanswered, clustered

Unanswered or unresolved visitor questions group by shared topic. Each cluster is a concrete writing task.

Scores

Beyond the average

A single score hides the failures that matter. Analytics shows where the assistant falls down, not just how often.

Share

Put the assistant where the questions are

Embeddable

A public chat

Share the assistant with visitors. Their names and emails - when they choose to leave them - collect locally in your browser.

No surprises

You do the follow-up

Nothing is emailed automatically. Export the contact list and follow up yourself, on your own email, on your own time.

Local-first

The boundary is visible

Sources, conversations, test cases, and visitor contacts live in your browser. DialogueBase sends nothing anywhere - the only way data leaves is when you export it yourself.