fix(indexer): resolve named embedder/VLM endpoints from the hydrated registry (fail loud on embedder mismatch)

[andyne13]

Background

#560 fixed the indexer Ray actor to hydrate the DB-backed model-endpoint registry and resolve named LLM endpoints (contextualization / topic tagging). The same root cause still affects other model types at index time, because the indexer's factories are built in IndexerPool.__init__ before the lazy registry hydration.

Problem

• Embedder — high severity (silent data corruption). _build_embedder_factory(cfg) returns None when cfg.models.embedder is empty at __init__ (it is, pre-hydration), so the factory is captured as None. _select_embedder then silently falls back to the global default embedder (cfg.embedder from env). A partition configured with a named embedder is silently ignored at index time → documents embedded with the wrong model → vectors incompatible with the query-side embedder → silently broken retrieval, with no error raised.
• VLM — lower severity. build_indexer_pool passes no vlm_factory (only a default VLM built from env cfg.vlm), so a per-partition named VLM is ignored for index-time image captioning, and admin-UI changes to the VLM endpoint never reach the indexer.

Not affected: the reranker (retrieval-only, runs in the API process) and the query/API process in general (it hydrates at startup and invalidates client caches on every endpoint CRUD).

Fix

• Give the embedder factory (and a new VLM factory) the same treatment the LLM factories got in fix(indexer): hydrate model-endpoint registry so admin-UI-registered LLM endpoints resolve #560: read the live cfg.models.* dict instead of capturing an empty snapshot / early-returning None at build time, and cache one client per endpoint identity (full-config hash, like _endpoint_identity) so endpoint edits propagate after a reload.
• Embedder must fail loud, never silent-fallback: an enabled named embedder that cannot be resolved must raise and fail the file — silently embedding with a different model (wrong vectors) is worse than a failed file. (This differs from the LLM enhancement path, which intentionally skips gracefully.)
• Wire a vlm_factory into the indexer pipeline for per-partition VLM selection.

References

Follows #560 / #554. Part of the v2.0 (refactor/hexagonal) release prep.

[coderabbitai]

🔗 Related PRs

#465 - fix(embedder): batch embedding requests to fix timeouts on big PDFs [merged]
#495 - fix(embedder): avoid vLLM exact-max_model_len hang; tune batch_size default [merged]
#511 - fix(indexer): keep pool actor alive for queued batches [merged]
#524 - fix(indexer): wire contextualization into hexagonal pool [merged]
#560 - fix(indexer): hydrate model-endpoint registry so admin-UI-registered LLM endpoints resolve [closed]

---

📝 Issue Planner

_{Check the box below or use the @coderabbitai plan command to generate an implementation plan and prompts that you can use with your favorite coding assistant.}

• Create Plan

---

🧪 Issue enrichment is currently in open beta.

You can configure auto-planning by selecting labels in the issue_enrichment configuration.

To disable automatic issue enrichment, add the following to your .coderabbit.yaml:

issue_enrichment:
  auto_enrich:
    enabled: false

💬 Have feedback or questions? Drop into our discord!

[hedhoud]

Closed by #564.