fix(core): L2-normalize FastEmbed vectors to satisfy unit-vector contract

[tk-pkm111]

Summary

• sqlite_search_repository.py converts vector distance to similarity with similarity = 1 - distance²/2, a formula that assumes unit-normalized vectors (the requirement is explicitly stated in comments on lines 65-67 of that file).
• Some embedding models — notably multilingual ones such as paraphrase-multilingual-MiniLM-L12-v2 — return vectors with norm ≈ 2.9 rather than 1.0. This causes every similarity score to collapse near 0.0, making vector/hybrid search effectively non-functional for those models.
• This PR applies L2 normalization inside _embed_batch() after converting raw model output to Python floats. Zero vectors are returned as-is to prevent division errors.

Root cause

FastEmbed's default English model (bge-small-en-v1.5) happens to return unit-normalized vectors, so the bug is invisible with the default config. Switching to any model that does not self-normalize (e.g., the multilingual MiniLM family) exposed the contract violation.

Changes

File	Change
src/basic_memory/repository/fastembed_provider.py	Add import math; L2-normalize each vector in _embed_batch()
tests/repository/test_fastembed_provider.py	Two new tests: norm==1.0 assertion for unnormalized stub model; zero-vector no-raise assertion

Test plan

• All 8 existing + new unit tests pass: uv run --python 3.12 pytest tests/repository/test_fastembed_provider.py -v
• New test test_fastembed_provider_l2_normalizes_output_vectors confirms output norm is 1.0 ± 1e-6 for a stub model that returns norm ≈ 2.9
• New test test_fastembed_provider_zero_vector_does_not_raise confirms no ZeroDivisionError on zero vector

🤖 Generated with Claude Code

[CLAassistant]

All committers have signed the CLA.

[phernandez]

Hi @tk-pkm111 — thanks for the careful fix and the clear reproduction. The change looks solid: the unit-norm contract on sqlite_search_repository.py lines 65-67 is real, multilingual MiniLM models do break it, and the patch addresses it at the right boundary. Tests look meaningful too.

The one blocker before we can merge is the CLA — the bot still shows it as unsigned for this PR. Could you sign it via the link in CLAassistant's comment above? Once that's in we can move toward landing this.

[tk-pkm111]

CLA is now signed — could you merge this when you get a chance?

[phernandez]

Reviewed the FastEmbed normalization change against the sqlite_search_repository unit-vector contract. No blockers found.\n\nLocal verification on PR head 4ef062b:\n- uv run pytest tests/repository/test_fastembed_provider.py -> 8 passed\n- uv run ruff check src/basic_memory/repository/fastembed_provider.py tests/repository/test_fastembed_provider.py -> passed\n\nNote: the repo's full Tests workflow is push-only and did not run for this external PR branch, so this review is based on focused local verification rather than the full CI matrix.