Move metadata extraction to an asynchronous pipeline #9

Closed
opened 2026-04-03 11:12:38 +00:00 by sam · 0 comments
Member

Summary

Refactor metadata extraction so captures do not depend on synchronous extraction during the request path, while still preserving eventual consistency for metadata quality.

Why

llm/plan.md lists async metadata extraction as deferred. AMCS already has retry and background-style recovery patterns; formalizing asynchronous extraction would improve latency and resilience under provider failures or slow upstreams.

Scope ideas

  • Persist the thought immediately with a pending metadata state
  • Queue or schedule extraction work outside the request-response path
  • Define retry/backoff and failure-state transitions
  • Ensure metadata status is queryable and observable
  • Keep existing retry tooling aligned with the new async flow
  • Document operational behavior and failure modes

Acceptance criteria

  • Thought capture succeeds without waiting for metadata extraction to finish
  • Metadata extraction runs asynchronously and updates stored thoughts later
  • Failure/retry behavior is defined and observable
  • Existing maintenance flows remain compatible or are replaced cleanly

Source

Deferred feature from llm/plan.md: async metadata extraction.

## Summary Refactor metadata extraction so captures do not depend on synchronous extraction during the request path, while still preserving eventual consistency for metadata quality. ## Why `llm/plan.md` lists async metadata extraction as deferred. AMCS already has retry and background-style recovery patterns; formalizing asynchronous extraction would improve latency and resilience under provider failures or slow upstreams. ## Scope ideas - Persist the thought immediately with a pending metadata state - Queue or schedule extraction work outside the request-response path - Define retry/backoff and failure-state transitions - Ensure metadata status is queryable and observable - Keep existing retry tooling aligned with the new async flow - Document operational behavior and failure modes ## Acceptance criteria - Thought capture succeeds without waiting for metadata extraction to finish - Metadata extraction runs asynchronously and updates stored thoughts later - Failure/retry behavior is defined and observable - Existing maintenance flows remain compatible or are replaced cleanly ## Source Deferred feature from `llm/plan.md`: async metadata extraction.
sam closed this issue 2026-04-03 11:27:12 +00:00
Sign in to join this conversation.