Skip to main content

Files Processing Pipeline

Every file that enters the workspace can run through the same chain: extract text → optionally categorize into collections → optionally ingest into the knowledge base. Uploads, cloud imports, folder sync, mailbox sync, and URL ingest all join that chain.

Your administrator chooses which steps run automatically. You still extract or ingest by hand from the Files grid whenever you need to.

The chain in plain language​

File arrives
  → extract text (and optional AI tags / excerpt)
    → propose or apply collections
      → ingest into the default vector index
StageWhat you see
ExtractText and metadata become available for search, preview, and AI
CategorizeCollection chips (or a proposal to accept)
IngestKnowledge column moves to processing, then completed

Ingest in the automatic chain needs collections and a default vector index. If categorization is in confirmation mode, ingest waits until you accept proposals.

Where files come from​

SourceTypical pathAdmin toggle
Manual or chunked uploadDrag-and-drop or BrowseAuto-extract on upload
One-shot cloud importImport from CloudAuto-extract on cloud import
Folder or mailbox syncSynced Folders / Synced MailboxesAuto-extract on sync
URL ingestURL ingestion in Files ConfigurationAuto-extract on URL

If auto-extract is off for a source, the file still stores — use Extract Content when you are ready.

Pipeline status​

Open Pipeline status from a folder (or the Files toolbar, depending on your plan). You see:

  • How many files still need extract, categorize, or ingest — counting only files that pass size, type, and organization guards
  • Files outside coverage (too large, excluded extension, missing organization when required)
  • For a synced folder, a Cloud sync card for files not imported yet

Corporate workspaces can Run configured pipeline on a folder to stamp extract → categorize → ingest using current settings. Developer accounts keep extract pending / re-extract / pause / cancel; they do not get the configured-pipeline button or the Files Configuration subtab.

You can pause processing for a folder path (a pipeline hold). That is separate from pausing a sync watch. A hold can last up to seven days. Canceling in-flight work can also set a hold until you resume.

Category proposals​

When smart categorization is on and set to confirm, the platform writes proposals instead of applying collections immediately.

  1. A banner or chips on the Files page indicate pending proposals
  2. Open the proposals dialog to accept or reject
  3. Accepted collections then allow chained ingest if auto-ingest is on

Re-running categorization on unchanged files can reuse cached results, so you may see no extra AI usage. That is expected.

Folder defaults​

A folder can override tenant defaults (collections, extract/ingest behavior) for files created under it. Open folder defaults from the folder menu. Ancestor folders are considered, so a nested folder inherits unless you override it.

Folder customer designation (Customers feature) also applies when a file is created — that writes customer links; it does not replace collections.

Authored taxonomy (collections)​

Administrators can import a closed set of collection labels (a taxonomy) so categorization matches your firm's vocabulary instead of inventing new tags. There is no in-app taxonomy editor: you import JSON, preview the diff, and apply. Deleting a collection in Collection Manager silently removes it from the closed set — the Taxonomy tab warns when a seeded label is missing.

Starter templates are available from Files Configuration → Taxonomy.

Best Practices​

  • Set a default vector index before turning on auto-ingest, or the chain stops after extract/categorize
  • Use confirmation mode for categorization until you trust the labels
  • Watch outside coverage in Pipeline status before assuming a backlog is stuck
  • For large synced folders, wait until discovery finishes before treating "remaining" as a complete count
  • Prefer Run configured pipeline on a folder over selecting thousands of rows for bulk extract

Troubleshooting​

Knowledge stays blank after upload​

Problem: Files appear but are not searchable in chat. Solution: Check extract status, then collections, then that a default index exists and auto-ingest is on. Accept pending proposals if confirmation mode is on.

Pipeline status numbers do not match the grid​

Problem: Pending extract looks smaller than the folder. Solution: Actionable counters exclude files that fail guards (size, extension, organization). Open Outside coverage for the rest. Synced files not yet imported appear on the Cloud sync card, not in extract totals.

Run configured pipeline is missing​

Problem: Developer Pipeline status has no that button. Solution: The remedy is corporate-only. Use Extract pending / re-extract instead, or ask an administrator on a corporate tenant.

Categorization keeps skipping​

Problem: Files stay uncategorized with no error. Solution: An empty model response is recorded as skipped and is not cached. Ask an administrator to check the categorization model setting. Very large files may also sit outside coverage.

⏱️ Read time: 12 minutes | 📊 Difficulty: intermediate | 🔄 Last updated: 2026-08-30