v1.2.0

Changed

• The project index embeds with EmbeddingGemma 2. The default model is now
  embeddinggemma-2:f32: Google DeepMind's google/embeddinggemma-2 at revision
  914f7f8, an open model under Apache-2.0 with no gate, text only, float32, on the CPU.
  It replaces nomic-embed-text. On a pre-registered benchmark of 93 studio queries over an
  11,525-chunk index it scored dense nDCG@10 +0.112 over nomic-embed-text, with a one-sided
  95% lower bound of +0.065 (code +0.174, prose +0.041). The 256-dimension cut lost 0.043,
  so vectors stay at 768. The record is in the rnd library, experiments/embeddinggemma-2/.
• Documents are sent with title: none | text: and queries with
  task: search result | query: . The code-retrieval prefix scored no better on code
  queries (−0.008), so one query prefix serves every kind.

Added

• tools/embed-serve, the CPU embed server. Ollama on Windows can't run EmbeddingGemma 2,
  so this small loopback server answers the part of Ollama's API that offrig's embed client
  uses, on the same default embed_url (http://127.0.0.1:11490).
  • It serves one name and answers 404 for any other, so no store can silently mix models.
  • It refuses to start when torch can see CUDA, and /api/ps reports size_vram: 0.
  • It encodes in sub-batches of 8.
  • It's built for a shared machine, after an outside review:
    • every thread pool (torch, OpenMP, MKL, OpenBLAS, the tokenizers) is capped before torch loads, by default to a quarter of the logical cores and at most 6 (--threads);
    • it runs at below-normal priority, single-threaded, one request at a time;
    • it bounds each request: 4 MiB, 256 texts, 12,000 characters a text, 2,048 tokens;
    • it refuses browser requests, non-loopback hosts, chunked or non-JSON bodies and a bad Content-Length.
  • Measured on a Core Ultra 9 285K with 6 threads on 6 efficiency cores: a 201-chunk rebuild took 163 s, the server used about 20% of the machine (never more than 22%), and peaked at 2.8 GiB. Its vectors match the benchmark's (cosine at least 0.9999999).
  • Python 3.12. requirements.txt pins every package exactly, with torch's CPU build.
  • Its tests use a stub encoder and run in CI's new embed-serve job, gated at 90% line
    coverage.
• Atlas: tools/embed-serve is its own part.

Upgrading: what a nomic-indexed store does

A project index built with nomic-embed-text keeps its vectors, and nothing mixes them with
the new model's:

• offrig index is refused: "the project index was built with nomic-embed-text, not
  embeddinggemma-2:f32; run offrig index --rebuild". It writes nothing.
• Hybrid search, the default once an index exists, embeds the query with the index's
  own model. The new server doesn't serve that model, so the search is refused, never
  downgraded. The refusal names offrig index --rebuild and mode=keyword.
• The side-car's background embed leaves new records keyword-only until the rebuild.

Re-embed each such project once, with embed-serve running:
offrig index --rebuild --project <dir>. Chunks are kept, and only vectors are replaced. A
project with no vectors yet needs nothing.

The refusal and start-line messages name embed-serve for the default model, and a CPU-only
Ollama for a model chosen with --model.