Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.teeem-ai.com/llms.txt

Use this file to discover all available pages before exploring further.

The Knowledge Base (KB) is a permanent store for your company’s documents. The agent draws from it when answering questions and shows the source. It’s not just embedding search — it has three-tier deduplication, version control, and sensitivity classification built in.
Teeem AI Knowledge Base management view

​
Concepts at a glance

Two sensitivity tiers

Public — available to everyone in the org Private — channel-scoped ACL, only authorised members

Three-tier dedup

Identical files are detected and skipped. Less indexing cost, less search noise.

Version chain

Newer versions soft-delete older ones; the change history is preserved.

Persistent citations

Sources used in an answer are written to a permanent message for audit purposes.

​
How to upload

Drag and drop files onto the chat input. The progress bar at the top of the chat shows status.

​
Sensitivity tiers

Sensitivity is auto-classified on upload and can be changed by an admin.
TierWho can search and cite
publicEveryone in the organisation
privateOnly members of the relevant channel/team (ACL-based)
When an item is promoted (public → private), the change is recorded in sensitivity_history so you can audit who changed what, when, and why.

​
Three-tier deduplication

Avoid indexing the same content twice.
1

Tier 1 — hash match

SHA256 of the file bytes plus a normalised text fingerprint. Identical content is skipped immediately.
2

Tier 2 — normalised fingerprint

A fingerprint over text with whitespace, line breaks, and page numbers stripped. Catches the same content shipped as a different format (e.g. PDF and DOCX of the same memo).
3

Tier 3 — vector similarity gate

For pairs above 0.92 cosine similarity, Gemini does a final semantic check. Only genuinely new content is indexed.
This three-tier gate typically cuts 25–40% of indexing cost.

​
Version control

Upload a new version of an existing document (e.g. policy.pdf → policy_v2.pdf) and the previous version is soft-deleted, replaced by the new one. Chunks from the old version are removed transactionally so they don’t surface in search.
  • supersede_reason — captures why a document was superseded (e.g. “policy revision”)
  • getDocumentVersionChain — admins can pull the full chronological version history
  • Version history panel — clicking a document shows previous versions, supersession reasons, and authors

​
Search and citations

When generating an answer, the KB runs:
  1. BM25 keyword search (Korean particle-stripping tokenizer)
  2. Vector semantic search (LanceDB)
  3. Reciprocal Rank Fusion to combine the two
  4. Re-ranking (optional, accuracy-first scenarios)

​
How citations show up

  • Slack — a context block with source cards beneath the answer
  • Web chat — citation cards alongside the answer (filename click previews the file)
  • Principle — citations are persistent messages, not inline footnotes — they’re easier to audit and harder to lose to edits

​
Supported formats

Excel, CSV, PDF, Korean HWP/HWPX, PPTX, DOCX, images (with OCR), video (with subtitles), audio, ZIP. See file processing for the complete list and per-format capabilities.