Each product has a document library. Documents are ingested: text extraction → chunking → embedding → indexing. The autoreply bot + AI generate use RAG search to cite relevant content.

​
Access

Products → :id → Documents tab (the product-doc table).

​
Supported file formats

FormatNotes
.pdfPage-by-page text extract
.docxWord document
.md / .txtMarkdown / plain text
.htmlTags stripped, text kept
.csvOne chunk per row
.png / .jpgLLM vision describes the image as text

​
Upload

1

Click + Upload (add-source modal)

Drag-drop or pick a file.
2

Metadata

  • Title (auto from filename, editable)
  • Description
3

Submit

File uploads to storage. Document marked status pending.
4

Background ingest

  1. Download the file
  2. Extract text per format
  3. Chunk (300–500 tokens, 50-token overlap)
  4. Embed each chunk via the Embed slot
  5. Save into the search index
  6. Status → indexed

​
Document status

StatusMeaning
pendingJust uploaded, awaiting ingest
processingWorker running
indexedReady for RAG
failedIngest failed
archivedSoft hidden, excluded from RAG

​
RAG search for the bot

When the autoreply bot receives an inbound, the system runs RAG search on the relevant product’s KB → top-K chunks are injected into the system prompt → the bot can cite them. See Autoreply.
Document detail tabs (Preview / Chunks / Usage / Settings), per-product manual search box, per-document or bulk re-index button, document visibility internal-only vs customer-facing — coming soon.

​
Best practices

Short documents, one topic per file → higher chunk precision.
Stale documents → the bot cites wrong info. Re-upload + archive the old version.