Research evidence.

CoHere FIRST1M: finite offered-load measurement

The recorded run uses 768-dimensional cosine vectors, top 10, and 64 sealed queries, repeated at each offered rate. Each call starts a fresh native process and scratch directory. Repeated rates do not add independent quality queries.

At 8 offered queries/s, all 64 calls succeeded with 96.71875% recall@10. Cold start to first HTTP response p90 was 541.10 ms. Completed full-span throughput, including drain, was 7.597044 queries/s.

Scheduled-to-response p90 was 542.36 ms, a different timing boundary. Sustained capacity, saturation, physical wire bytes, lifecycle billing, broader scale, and a matched vendor comparison are unmeasured in this run.

• Decision, protocol, and resource scope
• Independent root audit
• Measurement summary

Current native code and fixtures

The current path opens a small semantic directory, reads selected centroid leaves, shortlists two-bit SOURCE candidates, and scores authenticated SQ8 ranges. These source links pin the documentation audit's checkout; the measurement receipts record their own measured source.

• Semantic selection and leaf validation
• Generation open, admission, and search
• Generation and failure fixtures
• Logical IDs, mutations, recovery, and maintenance fixtures
• Delayed-delete fixtures

Archived research

The immutable research archive contains earlier experiments tied to their recorded source and configuration. Those reports do not define current capabilities or defaults.

For current public guidance, read the native API guide and maintenance coordination boundary. Python and TypeScript bindings remain separate experimental surfaces.