Skip to main content
The Teams Chat Compiler is the digestion layer between raw Microsoft Teams message ingestion and useful project intelligence. It reads Teams conversations from document_metadata, extracts structured business signals, and writes those signals back into the system.

​
What It Produces

After a Teams conversation is compiled, the system can have:
  • a plain-English overview of what happened;
  • tasks with owners, due dates, confidence, and review flags;
  • risks, decisions, sentiment, and initiative signals;
  • project attribution or attribution candidates for review;
  • a compiled status so the same conversation is not processed repeatedly.

​
Pipeline

The compiler runs each document through this sequence:
If a stage fails, the document should be marked with an error state that says which stage failed and why. Silent compiler failures are not acceptable because they make downstream AI answers look empty or unreliable.

​
Project Attribution

The compiler tries to connect each Teams conversation to the right project.
  1. Title override scans the chat title for project names, aliases, clients, and project numbers.
  2. Content inference searches the conversation text and participants when the title does not provide a strong match.
  3. Uncertain matches are written for human review instead of being silently auto-assigned.

​
Batch Behavior

Manual or scheduled runs process the newest conversations first. Low-content conversations are skipped so “ok,” “thanks,” or “sent” messages do not consume model calls or create noisy intelligence. Common run patterns:

​
Tables It Updates

  • document_metadata for overview, project attribution, tags, and status.
  • tasks for extracted action items.
  • insights for risks, decisions, sentiment, and initiative signals.
  • project_insights for high-confidence project-linked intelligence cards.
  • document_attribution_candidates for uncertain or competing project matches.

​
When To Check This Feature

Check the Teams Chat Compiler when:
  • Teams conversations are ingesting but executive briefs or project intelligence are missing the substance of those chats.
  • Documents remain in raw_ingested or embedded status for too long.
  • Project attribution looks wrong or incomplete.
  • Extracted tasks, risks, or decisions are too generic to trust.

​
Engineering Reference

The detailed implementation document lives at docs/archive/2026-06-22-docs-migration/features/teams-compiler.md.