fix(search): reserve interactive embeddings from their own admission lane

[waleedlatif1]

Summary

• Search queries embedded through the same per-credential admission bucket as bulk indexing, so during a crawl one interactive request competed with hundreds of batches
• Every embedding call still reserves from the credential's aggregate request and token buckets, so the configured budget is never exceeded. Checkpointed (bulk indexing) calls additionally reserve from bulk buckets capped at 90% of that budget, which leaves interactive callers guaranteed headroom
• Cooldown and quota gates stay shared, so a provider 429 or exhausted balance still pauses every caller

Type of Change

• Bug fix

Testing

• Admission tests assert bulk calls reserve aggregate plus 90% bulk buckets with unchanged cooldown keys, a bulk batch above the lane cap is rejected up front, and a one-per-minute budget still leaves one bulk request slot; client test asserts checkpointed calls are bulk and plain calls are not
• vitest on lib/embeddings, lib/core/rate-limiter, lib/knowledge/embeddings: 342 passing; type-check and lint 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)

🤖 Generated with Claude Code

[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 9:49am UTC

[greptile-apps]

Confidence Score: 5/5

The PR appears safe to merge; the current implementation fixes the previously reported aggregate-budget and undersized-bucket failures without introducing a new actionable defect.

Summary

• Bulk calls reserve from both aggregate buckets and bulk buckets capped at 90% of configured capacity.
• Oversized bulk token reservations are rejected before admission, while minimum request budgets retain one usable bulk slot.
• Embedding client tests verify checkpointed calls are classified as bulk and ordinary calls remain interactive.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[Embedding request] --> B[Reserve aggregate credential buckets]
    B --> C{Checkpointed bulk call?}
    C -->|No| D[Proceed using aggregate capacity]
    C -->|Yes| E[Reserve 90% bulk buckets]
    E --> F{Bulk token cost fits lane cap?}
    F -->|No| G[Reject before waiting]
    F -->|Yes| H[Proceed when aggregate and bulk capacity are available]
    D --> I[Shared cooldown and quota gates]
    H --> I

Loading

_{Reviews (3) · Last reviewed commit: "refactor(search): derive the bulk admiss..."}

[waleedlatif1]

@greptile review

[waleedlatif1]

@greptile review