fix(knowledge): wait for embedding admission instead of deferring documents for an hour

[waleedlatif1]

Summary

• Knowledge-path embedding batches now wait up to 60 s for the shared admission bucket instead of 5 s. Under load the expected queueing time is minutes, so a 5 s wait turned almost every contested batch into a full document re-dispatch with a 1-to-60-minute delay, for a limiter we run ourselves while the provider was healthy
• The embedding request bucket admits 64 concurrent starts instead of 8, so documents that begin together no longer lose a race for slots while the token budget sits unused. The per-minute rate is unchanged and still governs sustained throughput
• When an admission wait still expires, the document resumes after the bucket's stated wait, clamped to 10–60 s with jitter, and the yield counts against the processing-slice budget rather than the provider-failure attempts. Provider-side 429s and quota exhaustion keep the exponential ladder
• The admission deadline path now carries the bucket's last stated wait so the scheduler can use it

Type of Change

• Bug fix

Testing

• New tests: admission-timeout continuation uses the short jittered delay and slice budget, provider throttling keeps the exponential ladder, admission resumes are bounded independently, delay helper clamps and jitters; request burst asserted in admission tests; wait constant asserted in the embedding client test
• vitest across lib/knowledge/documents, lib/core/rate-limiter, lib/embeddings, background/knowledge-processing: 828 passing
• bun run type-check, bun run lint, bun run check:audits, docs-manifest:check pass

Checklist

• Code follows project style guidelines
• Self-reviewed my changes
• Tests added/updated and passing
• No new warnings introduced
• I confirm that I have read and agree to the terms outlined in the Contributor License Agreement (CLA)

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated
docs	Skipped		Sep 10, 2026 7:06am UTC

[greptile-apps]

Confidence Score: 5/5

The PR appears safe to merge; no new actionable failures or outstanding previous findings remain.

Summary

• Extends checkpointed embedding admission waits to 60 seconds while keeping them within the reserved request-retry budget.
• Raises the embedding request-bucket burst capacity without changing sustained per-minute throughput.
• Carries admission retry timing into short, jittered document continuations.
• Accounts admission timeouts against processing slices rather than provider-failure attempts.
• Adds focused coverage for admission bursts, retry-budget invariants, continuation accounting, and delay bounds.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart LR
  D[Document processing] --> C[Embedding client]
  C --> A[Shared provider admission bucket]
  A -->|Admitted| E[Embedding provider request]
  A -->|Wait expires| T[Admission timeout with remaining wait]
  T --> J[Clamp and jitter 10–60 seconds]
  J --> S[Dispatch processing-slice continuation]
  S --> D
  E -->|Provider throttling| P[Exponential provider continuation]
  P --> D

Loading

_{Reviews (2) · Last reviewed commit: "fix(knowledge): report only the admissio..."}

[waleedlatif1]

@greptile

[waleedlatif1]

@cubic-dev-ai review this PR

[cubic-dev-ai]

@cubic-dev-ai review this PR

@waleedlatif1 I have started the AI code review. It will take a few minutes to complete.