# Give Connect a document to answer from

If the document should shape answers everywhere and for months, it belongs in Knowledge. If it matters to one conversation or one deal, attach it there. The distinction is not about size — it is about whether you want the content retrievable by every future answer or scoped to the record it came with.

- **Status:** Available
- **Audience:** both
- **In the app:** #/knowledge, #/inbox
- **Last verified:** 2026-09-10
- **Canonical:** https://connectbyjbrh.com/docs/how-to/give-connect-a-document/

## Choosing between the two

| The document | Where it goes | Why |
|---|---|---|
| Service descriptions, policies, standard terms | Knowledge | Every channel's answers should be grounded in it |
| A price list you will be asked about | Knowledge | Commercial answers are refused unless the material supports them |
| One customer's signed order | Attachment on the deal | It answers questions about that deal, not about the business |
| A specification sent in an email thread | Attachment on the conversation | Scoped to the thread it arrived in |
| A spreadsheet you want to query | Attachment, or the Data grid | Reading a file is not the same as holding records |

> **Note** Loading a customer-specific document into Knowledge is the mistake worth avoiding. Knowledge grounds answers for everybody, so one customer's negotiated terms would start informing replies to other customers.

## Loading it into Knowledge

1. Open Knowledge and add the source.
   - Result: The file is read with the standard library: PDF via its text layer, DOCX, XLSX, CSV, TSV, PPTX, and plain text, Markdown or JSON. Images and a PDF without a text layer are read by the model instead.
2. Check what was extracted before you rely on it.
   - Result: Extraction is where a document quietly half-works. A scanned page and a table split across a page break are the two usual cases.
3. Review the facts drawn out of the source.
   - Result: A fact is one grounded statement. Facts are what a short answer quotes, and reviewing them catches a misread figure before a customer does.
4. Remove or correct anything the document says that is no longer true.
   - Result: Superseded material in Knowledge is worse than missing material, because an answer grounded in it looks authoritative.

> **Careful** A document from outside is untrusted text. Instructions embedded in it — a line telling a reader to ignore previous directions, for instance — arrive as ordinary business material and are handled as content rather than as commands. Read the prompt-injection page before loading documents you did not write.

## Proving the document is actually being used

Uploading a file and seeing it listed proves storage, not grounding. Three checks, in increasing strength.

1. **Ask the Assistant a question only that document answers.** If the reply cites the source, retrieval is working. If it answers vaguely without a citation, the content may be stored but not reachable.
2. **Run the retrieval test on the source.** This is the check built for exactly this question and it does not depend on how you phrased a prompt.
3. **Ask the same question on the channel that matters.** A phone call carries a tighter budget than a chat window — roughly two thousand characters of knowledge and about four facts — so a long document contributes less on a call than in an email.

That budget is the reason a document can work in the Assistant and appear to fail on a call. Nothing is broken; the call has room for less. Shorter, denser source material is worth more on the phone than a complete manual.

## When it does not take

**The upload was refused** — A declared DOCTYPE or entity, or an archive that expands far beyond its size, is refused rather than parsed.
**The file loaded but nothing was extracted** — A scanned PDF with no text layer. The model reads the image instead — check the result rather than assuming.
**Answers still say the material is not supported** — The claim needed is commercial, and commercial claims are refused unless Knowledge actually carries them. Add the figure explicitly rather than the surrounding prose.
**Two sources disagree** — The conflict is surfaced rather than resolved silently. Retire one of them; leaving both means every future answer inherits the same ambiguity.

## Questions

### Does a document in Knowledge get sent to a model for training?

No. Knowledge is retrieved to ground an answer at the time the answer is written. Connect does not train a model on your material.

### How much can I load?

There is no single number worth quoting, because what matters is not storage but how much fits into one answer. The per-answer budgets are what bind you, and they are tightest on a call, so curation beats volume.

### Can I give Connect a document during a conversation instead?

Yes — attach it to that conversation. It will be read for that thread, and it will not start influencing unrelated replies, which is usually what you want for something a customer just sent you.

## Related

- [Knowledge sources](https://connectbyjbrh.com/docs/knowledge/sources/)
- [The facts editor](https://connectbyjbrh.com/docs/knowledge/facts/)
- [Testing retrieval](https://connectbyjbrh.com/docs/knowledge/retrieval-test/)
- [Instructions hidden in content](https://connectbyjbrh.com/docs/knowledge/prompt-injection/)
- [Attach a document to a deal](https://connectbyjbrh.com/docs/how-to/attach-a-document-to-a-deal/)
- [File formats Connect reads](https://connectbyjbrh.com/docs/files-data/supported-formats/)
- [Knowledge and memory](https://connectbyjbrh.com/docs/compare/knowledge-vs-memory/)

## What this page is based on

- Connect source pack: overview (docs-source/sources/GENERAL.md §9 — files and formats)
- Connect capability registry (docs-source/facts.py — knowledge_sources, MEASURED knowledge budgets)
- Connect source pack: phone (docs-source/sources/PHONE.md §6 — what a session carries)