Give your AI agents memory of your codebase.

> Why did we move from Neo4j to Qdrant?

⏺ letapis · ena_get_context("Neo4j Qdrant migration decision")

  [decision] — 4 months ago · project: core
  Qdrant replaces Neo4j as the only storage engine.
  Why: vector search + payload filtering in one engine, simpler ops.
  source: IMPL_Qdrant_Migration.md

⏺ The team decided this in March: one engine for both vectors
  and payload filtering, and simpler ops. I'll follow that —
  details are in the migration doc.

> 

> Where do we split embedding batches when the token budget overflows?

⏺ letapis · search("embedding batcher token budget overflow flush")

  1. core/embedding/batcher.py · score 0.91
     def flush(self):  # splits the batch once the token budget is hit
  2. core/embedding/gateway.py · score 0.84
     …retry the oversized tail as its own batch…

⏺ Found it — batcher.py flushes on token budget, the oversized
  tail is retried by the gateway. No exact keyword needed:
  the query matched by meaning.

> 

> Same topic — but I want the design doc, not the code.

⏺ letapis · search("embedding batcher token budget",
      instruction="Find documentation explaining this concept")

  1. docs/embedding_pipeline.md · score 0.93
     "Batches flush on timeout or token budget, whichever first…"
  2. IMPL_Batcher_Rework.md · score 0.88
     rationale for the two-level flush + measured throughput

⏺ Same query, different intent: the instruction steers results
  toward docs that explain the design instead of the code
  that implements it.

>