# Connect said it does not know

Connect checked its Knowledge, found nothing with the authority to answer, and stopped before writing anything. That is the designed behaviour rather than a fault. What makes it actionable is the record left behind, which says what was retrieved, which of it carried authority, and which verdict the gate reached — three fields that usually name the fix.

- **Status:** Available
- **Audience:** both
- **In the app:** #/needs-you
- **Last verified:** 2026-09-10
- **Canonical:** https://connectbyjbrh.com/docs/troubleshooting/knowledge-insufficient/

## What you are looking at

On email: a thread in Needs You with a guidance request saying approved Knowledge does not contain a confident answer, and no draft beneath it. In the activity record for that message, a Knowledge check logged as a miss. On a call: the caller was told a colleague would come back to them, and a follow-up now exists for a person to honour.

The customer has received nothing. This is worth stating plainly because the queue can look like work in progress: it is not — it is a conversation that has stopped, with the customer's own clock still running.

## What it means

Retrieval and sufficiency are separate tests, and this is the second one failing. Material may well have been found; what was missing was material carrying the standing to settle the question. For an ordinary factual question that means a `verified_fact` or `owner_instruction` source, or a typed entry in an authority-carrying category. For a commitment — price, service level, warranty, timeline, support, anything legal, a specific meeting time — the bar is higher again: a topical policy will not do.

## The causes, most likely first

| Cause | How it shows |
|---|---|
| The right material carries only `reference` authority | Matched titles full, authoritative titles empty |
| The source is parked in `needs_review` | The document does not appear among matched titles at all, and its own record shows warnings or a conflict |
| Scope mismatch — the intent asked for scopes the source does not serve | Nothing matched, yet the same question finds it immediately under a different scope in search |
| A commitment topic needing explicit owner authority | A critical topic is listed, and the verdict names owner authority as required |
| Wording mismatch | A low score with unrelated titles — the customer's words appear nowhere in the material |
| Genuinely nothing on the topic | A score at or near zero and no titles |

> **Note** Work down this list rather than starting with the model. Five of the six are settings or content, and the sixth is a gap somebody has to fill; none of them is improved by regenerating.

## What Connect completed

- Received the message, canonicalised it into a thread and resolved the sender to a person.
- Ran the prompt-injection shield over the external text before any model call.
- Classified the intent and chose the scopes that follow from it.
- Retrieved and scored the eligible material, and judged its authority.
- Logged the check with the score, the titles, the authoritative titles, any critical topics, the verdict and the citations.
- Asked the workspace's owner for guidance, attaching the titles that nearly matched — or, if that request could not be delivered, held an empty draft carrying the reason.

## What Connect did not complete

- No reply was drafted. Drafting happens after the gate, so there is no text waiting to be approved.
- Nothing was sent, and the customer received no acknowledgement, holding message or apology.
- No memory was written about the question, and the thread was not closed.
- On a call, no factual answer was given — a call-back commitment was made instead, and that promise is now owed.

## What you can do now

1. Answer the guidance request in Needs You.
   - Result: That thread moves immediately, and the customer gets a reply written by a person or approved by one.
2. Turn your answer into a typed entry, with a category that carries authority and the customer's own words in it.
   - Result: The next identical question passes the gate. This is the single highest-value habit available here — a miss answered twice is a miss that was never fixed.
3. Search the same question and read which titles come back and under what authority.
   - Result: You are looking for the two-second diagnosis: right document, wrong authority; or nothing at all.

## What an administrator can do

- Clear the review queue. A source in `needs_review` answers nothing, and a workspace with three parked sources can feel entirely empty.
- Correct authority on documents that genuinely state current policy — deliberately, reading the re-check that follows.
- Widen a source's scopes, or add `all` where the material really is general.
- Split mixed documents so the parts that state facts carry fact authority and the marketing prose does not.
- Review how often commitment topics are firing. If most misses are pricing questions, the fix is a pricing policy on file rather than a change to the gate.

## When to escalate

Escalate when the record disagrees with the workspace: authoritative titles are present and the gate still reports a miss, or the check itself is failing rather than missing — a source that has become unreadable, or an integrity failure appearing repeatedly on a file nobody is touching. Those are not content problems.

Do not escalate a first miss on a topic nobody has ever written down. That is the system doing exactly what it should, and the answer to it is one entry.

## Questions

### Can I lower the threshold so answers come out more often?

The score a match must reach is configurable, and lowering it buys volume with accuracy — weak matches become confident answers. If misses are frequent, the cause is nearly always authority or scope rather than the number, and those can be fixed without spending accuracy.

### Why did an identical question get answered last week?

Check whether a source has moved into `needs_review` since — an integrity failure or a new conflict does that without anyone acting. Also check whether the message was classified as a different intent this time, which changes the scopes retrieval was allowed to use.

### Does this happen on WhatsApp and SMS too?

The gate is not a property of email. Any channel where Connect would state a business fact runs the same check, so the same miss appears with the same record behind it — and the same repair works for all of them at once.

## Related

- [Claims Knowledge does not support](https://connectbyjbrh.com/docs/knowledge/unsupported-claims/)
- [Testing retrieval](https://connectbyjbrh.com/docs/knowledge/retrieval-test/)
- [The facts editor](https://connectbyjbrh.com/docs/knowledge/facts/)
- [Knowledge sources](https://connectbyjbrh.com/docs/knowledge/sources/)
- [Troubleshooting](https://connectbyjbrh.com/docs/troubleshooting/)
- [Why uncertainty is a valid answer from a business agent](https://connectbyjbrh.com/research/uncertainty-is-an-answer/)

## What this page is based on

- `backend/app/runner.py` — the Knowledge gate, the guidance request and the held draft
- `backend/app/knowledge.py` — coverage, critical topics, authority reasons and the score setting
- Connect capability registry (docs-source/facts.py)