feat(memory_tree): switch embed model to bge-m3 (1024-dim, 8K context)

[sanil-23]

Summary

Migrates the Phase 4 embedder from nomic-embed-text (768-dim, 2048-token context) to bge-m3 (1024-dim, native 8192-token context).

Driver: nomic's hard 2048-token context cap was failing on long chunks. The chunker's char-based heuristic (DEFAULT_CHUNK_MAX_TOKENS = 3000) undercounts BERT-WordPiece tokens by ~1.5-2× for HTML-derived markdown — a 1500-chunker-token chunk routinely produces 2500+ real tokens, exceeding nomic's cap. Even with per-request num_ctx overrides, Ollama clamps to the model's GGUF-baked context_length, so the only fix is a model swap. bge-m3 has native 8192 context and is widely deployed for retrieval reranking.

⚠ Breaking (on-disk format)

mem_tree_chunks.embedding and mem_tree_summaries.embedding blobs from the 768-dim era are invalid against the new 1024-dim layout. unpack_embedding's post-call dim check will reject them on read. Operators upgrading need to either:

UPDATE mem_tree_chunks SET embedding = NULL;
UPDATE mem_tree_summaries SET embedding = NULL;

(workers re-embed on next access) — or wipe chunks.db entirely and re-sync.

The legacy fallback path in retrieval that drops Option::None embedding rows to the bottom of a semantic rerank handles the transitional state cleanly: stale rows still surface (via keyword/recency signals) until they're re-embedded.

Changes

File	Change
score/embed/mod.rs	EMBEDDING_DIM 768 → 1024; module + struct docstrings updated; test fixture updated
score/embed/ollama.rs	DEFAULT_MODEL nomic-embed-text → bge-m3; add EMBED_NUM_CTX = 8192 const; add num_ctx field to EmbedOptions request struct so Ollama doesn't clamp to its 4096 default; test assertions updated
score/embed/factory.rs	Test fixture switched to "bge-m3"

Test plan

• 26 unit tests pass in memory::tree::score::embed::*
• cargo check --bin openhuman-core clean

🤖 Generated with Claude Code

Summary by CodeRabbit

Release Notes

• Chores
  • Updated embedding model configuration and increased embedding dimensionality to improve semantic search accuracy.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro

Run ID: e49caf3d-0c7a-4c29-818e-b83aa53fe04d

📥 Commits

Reviewing files that changed from the base of the PR and between 644c5c8 and 984bbf7.

📒 Files selected for processing (3)

• src/openhuman/memory/tree/score/embed/factory.rs
• src/openhuman/memory/tree/score/embed/mod.rs
• src/openhuman/memory/tree/score/embed/ollama.rs

---

📝 Walkthrough

Walkthrough

The PR updates the default embedding model from nomic-embed-text to bge-m3 across three modules, increasing the embedding dimensionality constant from 768 to 1024, adding a num_ctx override option to Ollama requests, and updating corresponding tests and documentation.

Changes

Embedding Model Migration (nomic-embed-text → bge-m3)

Layer / File(s)	Summary
Data Shape src/openhuman/memory/tree/score/embed/mod.rs, src/openhuman/memory/tree/score/embed/ollama.rs	EMBEDDING_DIM constant updated from 768 to 1024. New EMBED_NUM_CTX constant (8192) added to Ollama module.
Core Implementation src/openhuman/memory/tree/score/embed/ollama.rs	DEFAULT_MODEL changed to "bge-m3". EmbedRequest now includes options field with EmbedOptions struct containing num_ctx override. Request construction updated to pass num_ctx: EMBED_NUM_CTX to Ollama.
Documentation src/openhuman/memory/tree/score/embed/mod.rs, src/openhuman/memory/tree/score/embed/ollama.rs	Module-level docs revised to reference bge-m3 model, native 8192 token context, and 1024 output dimensions. Migration notes added stating old 768-dimension embedding blobs must be wiped and re-embedded.
Integration & Tests src/openhuman/memory/tree/score/embed/factory.rs, src/openhuman/memory/tree/score/embed/mod.rs, src/openhuman/memory/tree/score/embed/ollama.rs	Factory test ollama_chosen_when_endpoint_and_model_set updated to use bge-m3. Unit tests unpack_wrong_dim_errors, happy_path_returns_embedding, and dim_mismatch_rejected updated to assert 1024 dimensions and bge-m3 model with num_ctx == 8192 in request payload.

Estimated code review effort

🎯 3 (Moderate) | ⏱️ ~20 minutes

Possibly related PRs

• feat: native embeddings module with Ollama and vector store #521: Directly related as both PRs modify the Ollama embedder implementation, update embedding dimensionality constants, and adjust corresponding test assertions.

Poem

🐰 A rabbit hops through embeddings bright,
From nomic texts to bge-m3's light,
Eight thousand tokens, one-oh-two-four dims tall,
Context flows deeper, we've answered the call! ✨

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 66.67% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title directly and clearly describes the main change: switching the embedding model from nomic-embed-text to bge-m3, including the key technical details (1024-dim, 8K context) that differentiate it.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

---

_{Review rate limit: 2/5 reviews remaining, refill in 27 minutes and 23 seconds.}

_{Comment @coderabbitai help to get the list of available commands and usage tips.}