embed_graph_tool returns ok but produces 0 embeddings when sentence-transformers is missing (silent fallback in get_provider)

[johndowbs]

Summary

embed_graph_tool returns {"status": "ok"} and reports nodes embedded, but list_graph_stats_tool continues to show Embeddings: 0 afterwards because the local embedding provider silently mis-reports availability when sentence-transformers is not installed. The tool should either succeed-and-actually-embed, or fail loudly with an actionable install hint. Right now it does neither — it returns "ok" and embeds nothing.

Version

code-review-graph==2.3.2 (also reproducible on main as of 4b6293f)

Environment

• Linux (Ubuntu 24.04, Python 3.12)
• sentence-transformers NOT installed (the package is from pip install code-review-graph without the [embeddings] extra)
• No CRG_OPENAI_*, GOOGLE_API_KEY, or MINIMAX_API_KEY env vars set
• Default provider (local)

Reproduction

1. pip install code-review-graph (no extras)
2. code-review-graph build --repo /some/repo — graph builds normally
3. Call embed_graph_tool — returns {"status": "ok", "newly_embedded": N, "total_embeddings": N} (or 0/0)
4. Call list_graph_stats_tool — reports Embeddings: 0 nodes embedded
5. graph.db mtime hasn't moved; no embed/sentence-transformers Python process was ever spawned

Root cause

In code_review_graph/embeddings.py:

class LocalEmbeddingProvider(EmbeddingProvider):
    def __init__(self, model_name: str | None = None) -> None:
        self._model_name = ...
        self._model = None  # Lazy-loaded   <-- import deferred

    def _get_model(self):
        if self._model is None:
            try:
                from sentence_transformers import SentenceTransformer
                ...
            except ImportError:
                raise ImportError(...)
        return self._model

__init__ does NOT import sentence_transformers. The import is deferred to _get_model().

In get_provider():

# Default: local
try:
    return LocalEmbeddingProvider(model_name=model)
except ImportError:
    return None

This try/except ImportError never fires because __init__ doesn't import anything. So get_provider() returns a non-None provider, EmbeddingStore.available = True, and embed_graph_tool's availability guard is bypassed.

Then in EmbeddingStore.embed_nodes, the deferred import would fire when self.provider.embed(texts) is called inside the if to_embed: branch — but here's the second issue: if to_embed is empty (which can happen if _node_to_text produces deduplicated text-hashes the first run finds zero novel entries, or if the table is somehow already partially populated from a previous half-finished run), the function returns 0 silently with no provider call ever made — and the user sees "ok, 0 embedded" with no diagnostic.

In our repro the more direct path was: embed_graph_tool reported success without writing any rows, and no Python child process was ever spawned (verified with ps/pidstat). That implies the ImportError did fire somewhere upstream and was swallowed, OR the to_embed list was empty due to a pre-existing schema row, OR embed raised an exception that was caught and returned as success.

Suggested fix

Make get_provider("local") actually verify availability:

# Default: local
try:
    import sentence_transformers  # noqa: F401
except ImportError:
    return None
try:
    return LocalEmbeddingProvider(model_name=model)
except ImportError:
    return None

(There's already a _check_available() helper in the file that does exactly this, but it's never called from get_provider().)

Once get_provider() correctly returns None when the local backend isn't installed, EmbeddingStore.available becomes False, and embed_graph_tool's existing guard fires the helpful error:

The local embedding provider needs sentence-transformers.
Install with: pip install code-review-graph[embeddings] —
or switch provider to 'openai' / 'google' / 'minimax'.

That's the desired UX.

Bonus diagnostic tool

A get_embedding_status_tool (or extending list_graph_stats_tool) to report:

• Currently configured provider (local / openai / google / minimax)
• Whether the provider is available (i.e., its backend imports succeed)
• The model name in use
• The Python interpreter / venv it would run in (since the MCP server can be run from various venvs and the provider's backend needs to be installed in the same one)

would help users diagnose this kind of mismatch quickly. We had to read source to figure out why "ok, 0 embedded" actually meant "no backend installed".

Acceptance criteria

• With sentence-transformers not installed and no cloud creds, embed_graph_tool returns status=error with the install hint, not status=ok with 0 embeddings.
• With sentence-transformers installed, embed_graph_tool actually embeds N>0 nodes and list_graph_stats_tool reflects the count.

[mvanhorn]

Submitted PR #484. Hoists _check_available() into get_provider so get_provider('local') returns None when sentence-transformers is missing instead of silently producing a no-op provider. 4 new tests cover both paths.

[tirth8205]

Shipped in v2.3.4 via release PR #512: https://github.com/tirth8205/code-review-graph/releases/tag/v2.3.4. Closing as completed. Note for #476: the default architecture overview output is now compact; full verbose output remains available explicitly with detail_level="standard".