Add offline Jina v5 nano retrieval embeddings

[RightL]

Summary

Adds --embedding-adapter jina-v5-nano for the merged jinaai/jina-embeddings-v5-text-nano-retrieval checkpoint. The Nemotron default and Jina v3.5 reranker remain available and unchanged.

• Load the local Safetensors snapshot and its official EuroBert code with local_files_only=True
• Use Query: / Document: prefixes, right-padding-aware last-token pooling, and FP32 L2-normalized output
• Use BF16 on CUDA or FP32 on CPU; report 768 dimensions and reject inputs beyond the 8,192-token context without truncating
• Include adapter settings and snapshot contents in cache identity, so changing embedding models rebuilds derived vectors
• Document model files, service launch, existing client configuration, health check, cache behavior, and CC BY-NC 4.0 license

This is a separate feature branch based on main after CI fix #12 (fa34141). It is for local evaluation and is not ready to merge on the strength of mocked tests alone.

Verification

GitHub Actions passed on Ubuntu and Windows for afeb9709a4449b19a2bd028096f7399b2db56206. Both ran the 9 new adapter tests with no skips. Full suite: 1,552 tests on each OS, 16 skips on Ubuntu and 44 on Windows; no failures/errors. All 7 native Windows lifecycle tests passed.

• python -m unittest tests.test_embedding_models tests.test_embedding_retrieval: 51 tests passed, including 9 new adapter tests
• python -m tests: 1,552 tests, 16 skips, 0 failures/errors on Linux
• python -m compileall -q rightmemory tests: passed
• git diff --check: passed

Adapter tests use dependency-free numeric doubles and mocked checkpoint loading. They cover offline loader arguments, CPU/CUDA precision choice, mixed-length pooling, prefixes, context boundaries, cache identity, CLI registration, service capability reporting, and the preserved Nemotron default. No model weights were downloaded. Actual Torch/Transformers checkpoint loading, GPU behavior, retrieval quality, latency, and VRAM usage remain to be tested on the model host.

Local evaluation

From a clean checkout:

git fetch origin
git switch --track origin/feat/jina-nano-embeddings
python -m pip install -e ".[embedding-service]"

Use your own downloaded snapshot of jinaai/jina-embeddings-v5-text-nano-retrieval, including model.safetensors, config.json, tokenizer.json, tokenizer_config.json, configuration_eurobert.py, and modeling_eurobert.py from the same revision. Point the service at the snapshot directory, not a GGUF/ONNX file or the base multi-task model. The existing downloaded Jina reranker v3.5 snapshot is still required.

rightmemory embedding-service --embedding-adapter jina-v5-nano --embedding-model "C:/models/jina-embeddings-v5-text-nano-retrieval" --reranker-model "C:/models/jina-reranker-v3.5" --device cuda:0 --port 8766

In a second terminal:

Invoke-RestMethod http://127.0.0.1:8766/info

Check that embedding_model starts with jina-v5-nano: and dimensions is 768. Keep or add this section in the Memory root configuration, preserving the existing agent/model fallback:

[retrieve]
backend = "embedding"
max_output_chars = 100000

[retrieve.embedding]
url = "http://127.0.0.1:8766"
candidate_count = 40
result_count = 10
timeout_seconds = 60

Then run your normal rightmemory retrieve --session <id> "<complete need>" command for the configured Memory root. First retrieval after switching models rebuilds the derived embedding cache. The service loads both models and does not download either one.