fix(inference,ai-settings): prevent + handle ollama embedding-model-as-chat (Sentry TAURI-RUST-4P6) (#3359)

[oxoxDev]

Summary

• A user picked an embedding model (bge-m3:latest) as their Ollama chat model → Ollama 400s "does not support chat" every turn → 36.6k Sentry events / 2 users on 0.57.13 (TAURI-RUST-4P6).
• Prevention (the real fix): hide embedding-only Ollama models from the chat/LLM model pickers so the misconfig can't be made.
• Handling (defense-in-depth): demote the 400 from Sentry (classifier) + replace the opaque JSON with an actionable "assign a chat-capable model" message.
• Capability comes from Ollama /api/show capabilities (already fetched for the context-window gate — no extra round-trip). Fail-open on unknown capability.

Problem

bge-m3:latest is OpenHuman's default memory-tree embedding model. Nothing stopped a user selecting it as their Ollama chat model, after which every chat turn failed:

ollama API error (400 Bad Request): {"error":{"message":"\"bge-m3:latest\" does not support chat","type":"invalid_request_error","param":null,"code":null}}

Two failures compounded it:

1. No prevention — the chat-model picker listed every installed Ollama model, embedding-only ones included.
2. Noise + opaque UX — the 400 bypassed the 404-only completion_only_404_guard, the classifier had no matching phrase (so it re-reported every retry → 36.6k events), and the user saw raw upstream JSON instead of remediation.

Deterministic user-state (capability mismatch), not a server bug — same family as TAURI-RUST-35 (does not support tools).

Solution

Layer 1 — prevention (root cause):

• Parse the Ollama /api/show capabilities list (previously deserialized but unused) and classify each model chat-capable / embedding-only / unknown (ollama_chat_capability, ollama.rs). Surface chat_capable on each diagnostics installed_models entry, fetched in the same /api/show round-trip that already resolves the context window.
• Frontend filters chat_capable === false out of the local-model pickers (CustomRoutingDialog + GlobalOwnModelSelector) via isChatSelectableLocalModel. Unknown (null) stays visible — fail-open. The embedding model is configured in a separate panel, so embedding selection is unaffected.

Layer 2 — Sentry demotion: add "does not support chat" to is_provider_config_rejection_message (config_rejection.rs). Demotes error→info via both the inline guard sites and api_error. 36.6k → 0.

Layer 3 — actionable error: not_chat_capable_guard (400/422) rewrites the opaque JSON into model '<m>' does not support chat — assign a chat-capable model in Settings → AI, preserving the phrase so it stays demoted. Wired into all chat-completions error paths.

Submission Checklist

If a section does not apply to this change, mark the item as N/A with a one-line reason. Do not delete items.

• Tests added or updated (happy path + at least one failure / edge case) — Rust: classifier verbatim body + 6 compatible guard/wire tests + ollama_chat_capability matrix; Frontend: 4 isChatSelectableLocalModel cases (false / true / unknown-fail-open / list-filter)
• Diff coverage ≥ 80% — filter+map extracted to pure toSelectableChatModels helper with unit tests so the changed FE lines are covered (Coverage Gate)
• Coverage matrix updated — N/A: behaviour-only change (no new feature row)
• All affected feature IDs from the matrix are listed — N/A: none touched
• No new external network dependencies introduced — wire tests use wiremock; capability uses the existing /api/show call
• Manual smoke checklist updated if this touches release-cut surfaces — N/A: error-classification + picker filter, no release-cut surface
• Linked issue closed via Closes #NNN in the ## Related section

Impact

• Desktop (all platforms). Touches the Ollama chat-model picker (hides embedding-only models) and the chat-completions error path.
• Security/perf: none. Capability reuses the existing /api/show round-trip — no extra requests.
• Observability: removes the 36.6k/2-user Sentry stream; event becomes an info-level config-rejection log.
• Compat / fail-open: older Ollama or an /api/show miss → chat_capable unknown → model stays selectable, so no usable model is ever hidden.

Related

• Closes: Embedding model picked as chat model floods Sentry with ollama 400 "does not support chat" (TAURI-RUST-4P6) #3359
• Sentry-Issue: TAURI-RUST-4P6 (self-hosted id 5338)
• Follow-up PR(s)/TODOs: none — prevention now covers the root cause.

---

AI Authored PR Metadata (required for Codex/Linear PRs)

Linear Issue

• Key: N/A
• URL: N/A

Commit & Branch

• Branch: fix/4p6-ollama-chat-capability-reject
• Commit SHA: f073111 (classifier) · 751fa71 (guard) · 7497730 (capability) · 3baf728 (picker filter)

Validation Run

• pnpm --filter openhuman-app format:check — prettier + rust fmt clean (verified locally)
• pnpm typecheck — clean
• Focused tests: cargo test --lib openhuman::inference::provider::{compatible,config_rejection} (151 + 11 ok), ...::local::ollama (44 ok); vitest run aiRouting.test.ts (10 ok)
• Rust fmt/check (if changed): cargo fmt + cargo check clean; cargo clippy --lib clean on changed files
• Tauri fmt/check (if changed): pnpm rust:check Finished clean (no shell change; verified post-submodule-init)

Validation Blocked

• command: N/A
• error: N/A
• impact: N/A

Behavior Changes

• Intended behavior change: embedding-only Ollama models no longer appear in chat-model pickers; the residual 400 is demoted + made actionable
• User-visible effect: can't misconfigure an embedding model as chat; clearer error if one slips through

Parity Contract

• Legacy behavior preserved: unknown-capability models stay visible (fail-open); embedding-model selection panel unchanged; all other 4xx/5xx handling unchanged; guard fires only on 400/422 carrying the exact phrase
• Guard/fallback/dispatch parity checks: not_chat_capable_ignores_unrelated_400 + not_chat_capable_requires_4xx_status; chat_capability_classifies_* covers completion/chat/embedding/unknown; api_error SessionExpired/classification on chat_with_history preserved (message upgraded post-hoc)

Duplicate / Superseded PR Handling

• Duplicate PR(s): none (searched open + merged for does not support chat / bge-m3 / 4P6 / config_rejection)
• Canonical PR: this
• Resolution: N/A

[coderabbitai]

📝 Walkthrough

Walkthrough

Adds fast-fail detection and actionable remediation for OpenAI-compatible backends (Ollama) rejecting embedding models used as chat models. Detects HTTP 400/422 "does not support chat" errors, replaces opaque JSON with actionable configuration guidance, wires guards into three chat request paths, and extends error classification to prevent Sentry re-reporting.

Changes

Embedding Model as Chat Model — Handling

Layer / File(s)	Summary
Core error detection and message generation src/openhuman/inference/provider/compatible.rs, src/openhuman/inference/provider/compatible_tests.rs	Adds is_not_chat_capable_model, not_chat_capable_model_message, and not_chat_capable_guard helpers to detect HTTP 400/422 errors containing "does not support chat", generate model-specific actionable remediation text, and wrap into fast-fail errors. Unit tests validate status-code requirements, signature matching, message enrichment, classifier integration, and guard firing.
Chat request path integration src/openhuman/inference/provider/compatible.rs, src/openhuman/inference/provider/compatible_tests.rs	Wires the guard into chat_with_system, chat_with_history, and native chat entry points to short-circuit on non-chat-capable model rejections. End-to-end test confirms chat_with_history fails fast with actionable message and preserves config-rejection classification without fallback.
Provider config-rejection classifier extension src/openhuman/inference/provider/config_rejection.rs	Extends is_provider_config_rejection_message to recognize "does not support chat" as a provider-config-rejection substring. Unit tests validate raw and enriched error bodies for TAURI-RUST-4P6 (embedding model as chat) classify correctly.

Estimated code review effort

🎯 3 (Moderate) | ⏱️ ~20 minutes

Possibly related PRs

• tinyhumansai/openhuman#3211: Adds fast-fail model capability mismatch guard in the same three chat entry points for completion-only model 404 rejection pattern.
• tinyhumansai/openhuman#2813: Extends is_provider_config_rejection_message to detect "does not support tools" rejection (same family of capability mismatch errors).
• tinyhumansai/openhuman#2346: Adds logging and suppression of provider config-rejection errors in streaming and native chat paths.

Suggested labels

sentry-traced-bug, bug, rust-core

Suggested reviewers

• M3gA-Mind
• senamakel

Poem

🐰 A rabbit hops through error logs with glee,
"Your embedding's confused—use chat, you'll see!"
Four hundred's now actionable, helpful, and true,
No more Sentry floods from a model askew. 🔧✨

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Linked Issues check	✅ Passed	All objectives from #3359 are met: the PR adds detection for 'does not support chat' to the config-rejection classifier, implements a guard generating actionable remediation messages, wires it into all chat completion paths, and includes comprehensive tests.
Out of Scope Changes check	✅ Passed	All changes are directly scoped to the linked issue: classifier updates, guard implementation, error path integration, and related test coverage; no unrelated modifications detected.
Docstring Coverage	✅ Passed	Docstring coverage is 100.00% which is sufficient. The required threshold is 80.00%.
Title check	✅ Passed	The title accurately summarizes the main changes: fixing Ollama embedding-model-as-chat handling (Sentry TAURI-RUST-4P6) by adding detection and actionable error messages.
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.

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

---

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

[M3gA-Mind]

Reviewed locally — well-structured, correctly layered fix (prevention + Sentry demotion + actionable error), with consistent fail-open behavior end to end and thorough tests. Approving in spirit; a few inline non-blocking notes below.

[M3gA-Mind]

Non-blocking: this does a raw substring match (does not support chat) over the whole sanitized body. Given the verbatim Ollama wire shape this is near-impossible to false-positive in practice, and the 4xx gate above + the not_chat_capable_ignores_unrelated_400 test guard it well. The only theoretical snag is an unrelated 400 whose body happens to contain that phrase in a different field (e.g. an error about some other feature not supporting "chat history"). No change needed — flagging for completeness.

[M3gA-Mind]

Worth making the coupling load-bearing in a comment here too (it already is in compatible.rs): this anchor phrase is what keeps the rewritten actionable message demoted. If anyone ever softens the rewrite in not_chat_capable_model_message and drops the phrase, the 36.6k stream reopens. The TAURI-RUST-4P6-enriched classifier test covers it — good — but a one-liner here pointing at that dependency would help the next editor.

[M3gA-Mind]

Nice — conservative in the right direction: only Some(false) when confident it's embedding-only, None everywhere ambiguous, and completion/chat wins over an embedding tag. The case/whitespace tolerance and the ["insert"]-only → None case are both covered by the matrix test. Mirrors the existing Unknown eligibility arm; consistent with the codebase's fail-open posture.

[M3gA-Mind]

Approving. Correct root-cause fix with well-reasoned defense-in-depth (prevention + Sentry demotion + actionable error), fail-open behavior that's consistent across Rust/TS/picker layers, exemplary documentation of the cross-module anchor-phrase coupling, and thorough tests (classifier matrix, guard signature, wiremock E2E, classifier-parity for both raw and enriched messages). The inline notes are non-blocking. Nice work.