fix(core): surface embedder config errors instead of silent mock fallback (#826)

[jyotirmya17]

Hi @RaghavChamadiya — thanks for the detailed breakdown on #826, it made the fix straightforward to scope correctly. This follows the architecture you suggested to resolve the conflation issue without falling back to clamping.

Summary

• Shared validation helper: Added parse_numeric_env() and a typed EmbedderConfigError in core/providers/embedding/base.py. Strictly validates numeric inputs (rejects NaN, Inf, and values <= 0), shared across both dimensions and timeout.
• CLI no longer masks config errors: build_embedder() now raises EmbedderConfigError explicitly instead of silently falling back to MockEmbedder. This is intentional — a bad config should crash the CLI, not degrade quietly.
• MCP degradation path unaffected, now more precise: _server.py's existing broad try...except Exception around embedder instantiation already catches this, so EmbedderConfigError bubbles into the degradation envelope naturally — _embedder_status["reason"] now reports the exact config error instead of a generic failure.
• Constructors kept in sync: Both _embedder_kwargs() and OllamaEmbedder.__init__() now go through the shared validator, so validation logic isn't duplicated or drifting between the two.

Scope note: No changes to public constructor signatures — only internal validation logic. Existing valid configs are unaffected; only invalid ones (previously silently mocked) now fail loudly.

Related Issues

Resolves #826

Test Plan

• pytest passes — added 4 unit tests covering EmbedderConfigError on NaN, Inf, zero, and negative inputs for both dimensions and timeout
• ruff check . passes
• All existing tests still pass, no regressions

Checklist

• Follows project code style
• Tests added for new functionality
• Existing tests pass
• Docs updated where relevant (n/a — no public API surface changed)

[repowise-bot]

✅ Health: 7.6 (unchanged)

📋 At a glance
4 hotspots touched · 1 file with recent fix history.

Files & modules (2)

• packages (2 files)
  • .../providers/embedders.py
  • .../embedding/ollama.py
• tests (2 files)
  • .../cli/test_shared_helpers.py
  • .../test_persistence/test_ollama_embedder.py

🩹 Review priority _{(files here with the most recent bug-fix history — defects cluster, so review these first)}

• .../providers/embedders.py — fixed once in the last ~6 months

🔎 More signals (1)

🔥 Hotspots touched (4)

• .../providers/embedders.py — 4 commits/90d, 6 dependents · primary owner: biggiesmallcap-blip (40%)
• .../cli/test_shared_helpers.py — 3 commits/90d, 0 dependents · primary owner: Swati Ahuja (82%)
• .../embedding/ollama.py — 2 commits/90d, 5 dependents · primary owner: biggiesmallcap-blip (89%)

1 more

• .../test_persistence/test_ollama_embedder.py — 2 commits/90d, 0 dependents · primary owner: biggiesmallcap-blip (79%)

👀 Suggested reviewers @biggiesmallcap-blip

---

📊 Full report · ⭐ Star Repowise · 📥 Install bot · Last updated 2026-07-15 05:58 UTC
_{Silence on a single PR with [skip repowise] in the title · Per-repo toggle on repowise.dev/settings?tab=bot}

[RaghavChamadiya]

Thanks for turning this around quickly. The shared helper and typed error are the right shape, and I'm glad you raised rather than clamped.

One blocking item, on point 2 from my issue comment.

The config error is still swallowed on the search path. build_embedder's fallback wasn't the only swallow; the call sites have their own. search_cmd.py:166 and :286 wrap it in try: ... except Exception: pass and fall through to FTS, and update_cmd/incremental.py:31 does except Exception: return None. Against this branch:

OLLAMA_EMBEDDING_TIMEOUT=abc, repowise search
build_embedder RAISED: EmbedderConfigError -> Invalid OLLAMA_EMBEDDING_TIMEOUT: 'abc'
search_cmd guard: except Exception: pass
=> swallowed, silent fallthrough to FTS

So the reported symptom is unchanged; the swallow moved up a level. This is what the _embedder_status suggestion was reaching for: an envelope reports the error even with a broad except between the raise and the user. Raising alone can't survive those call sites. Relatedly, reindex_cmd.py:63 calls it bare and now shows a raw traceback, bypassing the MockEmbedder check plus click.Abort() right below it, which is the pattern worth extending.

MCP gemini mapper still has the original bug. _server.py:68 has its own _embedder_kwargs doing int(dims) if dims else 768, untouched, so REPOWISE_EMBEDDING_DIMS=-5 reaches the constructor unvalidated. The Ollama half of the MCP claim is right; this second copy needs the validator too.

Smaller: the error names the wrong var when value is picked by truthiness but name by presence (OLLAMA_EMBEDDING_DIMS="" + REPOWISE_EMBEDDING_DIMS=-5 reports the former); OllamaEmbedder(timeout=-5) skips validation via the is not None branch; the except EmbedderConfigError: raise block is a no-op; B904 needs from None. And list_issues.py looks like a debug script committed by accident, please drop it.

Tests pass (36 green). The body mentions NaN/Inf/zero tests but the four cover "invalid" and "-5" only; the validator handles the rest correctly, so it's just a coverage gap.

Happy to look again once config errors thread through an envelope instead of raising into call sites that catch broad Exception.

[jyotirmya17]

Thanks for the thorough review, Raghav! That makes complete sense, I completely missed that those upstream call sites had broad exception catchers swallowing the new typed error.

I'll refactor this to use the envelope pattern/status object so config errors can safely bubble past those try/except blocks and reach the user. I'll also update the MCP gemini mapper to use the validator, patch up the smaller logic bugs, drop the accidental list_issues.py script, and fill in the missing test coverage.

I'll get these changes done as soon as possible and push an update to the branch.