fix(inference): route vision at vision-capable models and refresh local model defaults (#5146)

[M3gA-Mind]

Summary

• Vision requests could be routed at models that cannot see. MVP_ALLOWED_VISION_MODELS was &[""], so every configured vision model was rewritten to the empty string and handed to Ollama as a nameless model. Replaced with a real capability check.
• New inference::vision_models registry encoding which local model IDs actually accept image input, so a vision tool-call reaches a vision-capable model or surfaces a clear error, never a chat-only model.
• DEFAULT_OLLAMA_VISION_MODEL was ""; it is now moondream:1.8b-v2-q4_K_S, the smallest genuinely vision-capable model that pulls with no extra setup.
• Default model list audit (§1.3). Two preset chat models were missing from the allowlist (so those tiers silently downgraded to the 1B model), and the 16 GB+ tier used a text-only model for vision.
• New user docs covering local (Ollama) and BYOK provider setup, and what each supports for chat / vision / embeddings.

Problem

1. Vision resolution was structurally broken. MVP_ALLOWED_VISION_MODELS: &[&str] = &[""] is an allowlist whose only member is the empty string, so effective_vision_model_id redirected every configured vision model to "", including genuinely vision-capable ones. A user who set local_ai.vision_model_id = "llava:7b" on the Custom tier passed the VisionMode::Disabled guard (non-empty id implies Ondemand), then reached ensure_ollama_model_available(config, "", "vision"), which issued POST /api/pull {"name": ""} and retried it three times with backoff before failing with an opaque error.

2. A vision tier pointed at a blind model. The 16 GB+ preset used gemma3n:e4b-it-q8_0 as its vision_model_id with VisionMode::Bundled. Gemma 3n is text-only on Ollama. This fails quietly rather than loudly: Ollama accepts an images array against a text-only model, discards it, and answers from the prompt text, so the user receives a fluent and entirely fabricated description of an image the model never saw.

3. Presets disagreed with the allowlist. gemma3:270m-it-qat (1 GB tier) and gemma3:4b-it-qat (8-16 GB tier) were not in MVP_ALLOWED_CHAT_MODELS, so applying either preset resolved straight back to gemma3:1b-it-qat. The user picked a tier and silently got a different model than the one advertised.

Solution

Vision capability registry

src/openhuman/inference/vision_models.rs maps model IDs to image-input capability, keyed on whole family names rather than loose substrings so near-misses cannot collide. It encodes the two traps this bug hit:

• Gemma 3 is split by size. 270M and 1B are text-only; 4B/12B/27B are multimodal. gemma3:latest is the 4B build.
• gemma3n is not gemma3. It is a separate, text-only model that shares the prefix. It is the load-bearing entry in TEXT_ONLY_FAMILIES.

Unknown IDs resolve to "not vision-capable", so the caller reports a missing vision model rather than shipping images to a model that will ignore them.

Resolution path

• enforce_vision_capability replaces enforce_mvp_vision_allowlist. The tier restriction the old allowlist stood in for is already enforced upstream by presets::vision_mode_for_config, which reports VisionMode::Disabled for tiers that ship no vision model; what remained was purely the capability question. A chat-only configured model now logs a warning and falls back to a vision-capable default rather than being passed through.
• effective_vision_model_id keeps its String signature for status/reporting surfaces (empty means "not configured", which is a legitimate state), but a non-empty return is now always vision-capable.
• New resolve_vision_model_id is used by vision_prompt for actual requests. It never returns an empty ID: an unconfigured vision model produces a message naming the config key to set and the models to pull. A configured-but-unpullable model now reads as a vision problem with the exact ollama pull command, rather than a generic pull failure.

Default model list audit (§1.3)

Every ID below was verified against the live registry, GET https://registry.ollama.ai/v2/library/<name>/manifests/<tag> returning 200, plus the published capability badges on ollama.com/library/<name>.

Change	Reason
MVP_ALLOWED_CHAT_MODELS gains gemma3:270m-it-qat, gemma3:4b-it-qat	Preset chat models that were being silently downgraded
16 GB+ tier: gemma3n:e4b-it-q8_0 to gemma4:e4b-it-q8_0 (chat + vision)	Gemma 3n is text-only; Gemma 4 is multimodal at every size
8-16 GB and 16 GB+ embeddings: nomic-embed-text:latest to bge-m3	Behaviour-neutral: the embedding allowlist already rewrote it to bge-m3 at resolution time. Memory Tree needs 1024-dim vectors
approx_download_gb refreshed on both tiers	Match what is actually pulled

On re-adopting gemma4: #5055 moved this tier off gemma4:e4b because no gemma4 namespace existed on Ollama at the time. That is no longer true. Gemma 4 has since been published, gemma4:e4b-it-q8_0 resolves (11.6 GB, 128K context), and it is multimodal, so the tier returns to one model serving both chat and vision, matching how the 8-16 GB tier uses gemma3:4b-it-qat. gemma3n:e4b-it-q8_0 stays allowlisted for back-compat with anyone who already pulled it.

Documentation (Part 3)

• New gitbooks/features/model-routing/local-and-byok-models.md: the three routes (managed / BYOK / local) with a capability comparison, per-model chat-vision-embeddings table, RAM tier table, provider slugs, workload routing fields, mixing routes, and troubleshooting keyed to the new error messages.
• local-ai.md: new local-vision section explaining why a chat-only model fails silently, plus corrected embedding default and download sizes.
• model-routing/README.md and the root README now state plainly that the subscription is a default rather than a requirement.

Submission Checklist

• Tests added or updated (happy path + at least one failure / edge case)
• Diff coverage ≥ 80% — enforced by the coverage gate in CI, which is the verification path for this PR (local runs were disallowed for this task, see Validation Blocked). The vision_prompt code paths changed here are now driven end-to-end against a mock Ollama server rather than only through their resolution helpers.
• N/A: behaviour-only change — no feature rows added, removed, or renamed
• N/A — no new matrix feature IDs
• No new external network dependencies introduced — registry lookups were manual verification during development, not code
• N/A: no release-cut surface touched — local model defaults and docs only
• Linked issue referenced in ## Related

Impact

• Runtime: Rust core (desktop + headless). No UI changes, therefore no i18n changes — zero files touched under app/.
• Behaviour change, intended: local vision now works when configured, instead of being silently disabled. Users on the Custom tier with a valid vision_model_id will start actually pulling and using that model.
• Compatibility: current_tier_from_config matches on model IDs, so users already on the 8-16 GB or 16 GB+ tiers will reverse-lookup as Custom after this change. Both tiers are outside the current MVP ceiling (MVP_MAX_TIER = Ram2To4Gb) and so are not reachable through the normal preset flow.
• Download size: 16 GB+ tier grows from ~9.9 GB to ~12.8 GB because Gemma 4 q8_0 is larger than Gemma 3n q8_0, and it replaces a separate vision model with one multimodal model.
• Security/migration: none.

Related

• Closes:
• Part of Ollama + local model full audit: install, test all model categories end-to-end, fix provider errors, spurious API key prompts, and vision — with documentation #5146 (§Part 1 vision, §1.3 default model list audit, §Part 3 documentation). Sibling streams cover the other sub-parts, so this PR intentionally does not close the issue.
• Context: supersedes the expired gemma4-does-not-exist assumption from Fix local model reliability: validate local model behavior in Tiny Agents and correct provider configuration #5055.
• Follow-up PR(s)/TODOs: the 1 GB and 4-8 GB tiers still use all-minilm:latest (384-dim), which the Memory Tree's 1024-dim on-disk format rejects at embed time. Left alone here because it is allowlisted for back-compat and is an embeddings concern rather than a vision one.

---

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

Linear Issue

• Key: N/A
• URL: N/A

Commit & Branch

• Branch: feat/GH-5146-vision-model-defaults-docs
• Commit SHA: 131c7f9

Validation Run

• pnpm --filter openhuman-app format:check — N/A, no frontend files changed
• pnpm typecheck — N/A, no TypeScript changed
• Focused tests: run in CI via Rust Core Coverage (libtest filter openhuman::inference). All 18 tests added in the first commit passed on the first run; the three vision_prompt mock-server tests were added in the second.
• Rust fmt: cargo fmt run and clean (formatter only, no compilation)
• Tauri fmt/check: N/A, app/src-tauri untouched

Validation Blocked

• command: cargo check / cargo test / any compile step
• error: not run — explicitly disallowed for this task; CI is the verification path
• impact: compilation and test results are unverified locally. Update: CI has since compiled and run everything. Rust Quality (fmt, clippy) and Rust Feature-Gate Smoke (gates off) passed, and all 18 new tests passed. The one failure was an unrelated pre-existing race in install_piper, fixed in the second commit and explained below. Original note follows: To compensate, every consumer of the changed symbols was manually audited (effective_vision_model_id call sites in bootstrap.rs, assets.rs, model_pull.rs, diagnostics.rs, types.rs — all are gated behind a non-Disabled vision mode or are report-only), test names were checked for collisions across the inline and sibling test modules in presets.rs, and no existing test was found to pin the preset values that changed. Please treat CI as the gate on this PR.

Behavior Changes

• Intended behavior change: vision requests resolve to a vision-capable model or fail with an actionable message; they can no longer be routed at a chat-only model or at an empty model name.
• User-visible effect: local vision works when configured rather than silently doing nothing; clearer errors when it is not configured or not pulled; the 16 GB+ tier gains genuine vision.

Parity Contract

• Legacy behavior preserved: effective_vision_model_id keeps its signature and its empty-means-not-configured contract; gemma3n:e4b-it-q8_0 stays allowlisted; the LM Studio bypasses on chat and embeddings are untouched; the embedding allowlist is unchanged.
• Guard/fallback/dispatch parity checks: vision_mode_for_config remains the tier gate, unchanged. Every existing ensure_ollama_model_available vision call site was verified to sit behind a non-Disabled vision mode. Both-directions tests pin the new invariants (preset vision models are vision-capable; preset chat models resolve unchanged).

CI Fix (second commit)

non_executable_workspace_binary_is_skipped_so_path_can_win failed with
NotFound on its second set_permissions. It is not related to the vision
change: workspace_piper_binary_candidates resolves through
paths::shared_root_dir, which ignores config.workspace_dir unless
OPENHUMAN_WORKSPACE is set and otherwise returns the real
~/.openhuman/bin/piper. The test's temp_config() TempDir therefore never
isolated it, and it was the only test in the module writing to that shared
directory without taking shared_install_lock(), while siblings call
wipe_shared_install_dir.

A latent race, surfaced rather than caused by this PR: CI Lite derives the
libtest filter from changed paths, so touching src/openhuman/inference/...
selects openhuman::inference, and the tests added here changed the parallel
scheduling enough to interleave the write and the wipe. Fixed by following the
convention already used by every other test in that file.

Duplicate / Superseded PR Handling

• Duplicate PR(s): none
• Canonical PR: this one
• Resolution: N/A

Summary by CodeRabbit

• New Features
  • Added and documented three inference routes (managed, bring-your-own-key, fully local) with per-workload routing.
• Improvements
  • Updated local vision handling to use capability checks, preventing image requests from going to chat-only models.
  • Refreshed defaults across presets (including bge-m3 embeddings and gemma4 multimodal wiring) and improved Moondream-based vision configuration.
  • More actionable error messaging when vision is missing or models need to be pulled/rerouted.
• Documentation
  • Updated model-routing and local-AI guides, added a new “Local models & bring your own key” page, and expanded related links.

[M3gA-Mind]

@coderabbitai review

[coderabbitai]

✅ Action performed

Review finished.

Note: CodeRabbit is an incremental review system and does not re-review already reviewed commits. This command is applicable only when automatic reviews are paused.

[coderabbitai]

Important

Review skipped

No new commits to review since the last review.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro

Run ID: 6aeb2eaf-73ef-4f5a-9c03-aa9c2fb0572b

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

📝 Walkthrough

Walkthrough

The change adds capability-based local vision model resolution, updates local model presets and embeddings, isolates installer test state, and documents managed, BYOK, and fully local routing options.

Changes

Model routing and local vision

Layer / File(s)	Summary
Vision capability registry and resolution src/openhuman/inference/...	Adds centralized vision capability detection, new defaults and resolution errors, refreshed model allowlists, and expanded tests.
Vision request failure handling src/openhuman/inference/local/service/vision_embed.rs	Validates vision models, records missing state, and returns actionable errors for unavailable or invalid models.
Preset model and invariant updates src/openhuman/inference/presets.rs	Updates RAM-tier models, download estimates, and validation for qualified and vision-capable models.
Local and BYOK routing documentation README.md, gitbooks/...	Documents local, managed, and BYOK routes, configuration, vision behavior, troubleshooting, and navigation links.
Shared installer test isolation src/openhuman/inference/local/install_piper.rs	Isolates workspace and shared Piper install state during tests.

Estimated code review effort: 4 (Complex) | ~45 minutes

Sequence Diagram(s)

sequenceDiagram
  participant VisionRequest
  participant ModelResolver
  participant Ollama
  VisionRequest->>ModelResolver: Resolve vision-capable model
  ModelResolver-->>VisionRequest: Return model ID or configuration error
  VisionRequest->>Ollama: Ensure model availability and generate response
  Ollama-->>VisionRequest: Return result or unavailable-model error

Loading

Suggested labels: bug, rust-core

Poem

I’m a rabbit with models to choose,
Moondream eyes and embeddings to use.
Local or cloud, the routes align,
BYOK hops neatly through the design.
If vision goes missing, guidance appears.

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly matches the main changes: vision routing now targets vision-capable models and local model defaults were refreshed.
Docstring Coverage	✅ Passed	Docstring coverage is 100.00% which is sufficient. The required threshold is 80.00%.
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.

---

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

[greptile-apps]

Greptile Summary

This PR fixes three distinct bugs in local vision model routing: (1) MVP_ALLOWED_VISION_MODELS = &[""] was an allowlist whose only member was the empty string, silently rewriting every configured vision model — including genuinely capable ones — to "", then retrying a nameless POST /api/pull three times before failing opaquely; (2) the 16 GB+ preset used gemma3n:e4b-it-q8_0 as its vision model, a text-only model that Ollama accepts without error but answers from prompt text alone; (3) two preset chat models were absent from the chat allowlist, silently downgrading users to the 1B default.

• New vision_models.rs registry classifies model IDs as vision-capable or chat-only using exact family matching, with special-case size-aware logic for gemma3 (4B+ multimodal, smaller text-only) and an explicit TEXT_ONLY_FAMILIES denylist preventing gemma3n from ever matching the gemma3 family rule.
• resolve_vision_model_choice replaces effective_vision_model_id on the actual-request path, guaranteeing a non-empty vision-capable ID or an actionable error; the new VisionModelChoice.replaced field surfaces any capability substitution directly in the pull-failure error message.
• Preset and allowlist audit: gemma3:270m-it-qat and gemma3:4b-it-qat added to MVP_ALLOWED_CHAT_MODELS; 16 GB+ tier updated to gemma4:e4b-it-q8_0; embedding defaults corrected to bge-m3; SharedRootOverride guard fixes a latent test race in install_piper.

Confidence Score: 5/5

Safe to merge. All three core bugs are correctly fixed, the capability registry errs conservatively toward false for unknown models, and the substitution reporting is well-covered by tests.

The vision routing logic is sound end-to-end: empty-string allowlist is gone, every path through resolve_vision_model_choice either returns a verified vision-capable ID or a clear actionable error, and the gemma3/gemma3n split is pinned by dedicated regression tests. No correctness or behavioral regressions were found.

Files Needing Attention: No files require special attention. The VISION_MARKERS substring fallback in vision_models.rs is a deliberate heuristic for repackaged upstream models and is documented as such.

Important Files Changed

Filename	Overview
src/openhuman/inference/vision_models.rs	New capability registry; uses exact family matching with TEXT_ONLY_FAMILIES denylist and size-aware gemma3 logic. Conservative (unknown = false). Well-tested including regression guards for gemma3n.
src/openhuman/inference/model_ids.rs	Replaces empty-string vision allowlist with enforce_vision_capability; adds resolve_vision_model_choice with VisionModelChoice.replaced for transparent capability-substitution reporting; expands MVP_ALLOWED_CHAT_MODELS to cover all preset models.
src/openhuman/inference/local/service/vision_embed.rs	Uses resolve_vision_model_choice instead of effective_vision_model_id; surfaces substitution_note in pull-failure errors; three new mock-server tests drive the full vision_prompt path end-to-end.
src/openhuman/inference/presets.rs	8-16 GB tier embedding corrected from nomic-embed-text to bge-m3; 16 GB+ tier chat/vision updated to gemma4:e4b-it-q8_0; download sizes updated; old gemma4-namespace assertion replaced with durable invariants.
src/openhuman/inference/local/install_piper.rs	Adds SharedRootOverride RAII guard to redirect OPENHUMAN_WORKSPACE to TempDir for the duration of the permissions test, fixing a latent race that parallel test scheduling exposed.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[vision_prompt called] --> B{vision_mode_for_config}
    B -- Disabled --> C[Return early: vision disabled]
    B -- Ondemand / Bundled --> D[resolve_vision_model_choice]
    D --> E{effective_vision_model_id}
    E --> F{config.vision_model_id empty?}
    F -- Yes --> G[Return Err: no vision model configured]
    F -- No --> H{Alias normalization?}
    H -- Yes --> I[Normalize to DEFAULT_LOW_VISION_MODEL]
    H -- No --> J[Use raw value]
    I --> K[enforce_vision_capability]
    J --> K
    K --> L{is_vision_capable?}
    L -- Yes --> M[Return id unchanged]
    L -- No --> N[Warn + return DEFAULT_OLLAMA_VISION_MODEL]
    M --> O[VisionModelChoice: model=id, replaced=None]
    N --> P[VisionModelChoice: model=moondream, replaced=Some configured id]
    O --> Q[ensure_ollama_model_available]
    P --> Q
    Q -- Ok --> R[Issue vision request with images array]
    Q -- Err --> S[Return Err: vision model unavailable + substitution note]

Loading

_{Reviews (5): Last reviewed commit: "fix(test): drop the unreachable vision r..." | Re-trigger Greptile}

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 131c7f9931

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[coderabbitai]

Actionable comments posted: 3

Caution

Some comments are outside the diff and can’t be posted inline due to platform limitations.

⚠️ Outside diff range comments (1)

src/openhuman/inference/presets.rs (1)

160-211: 🎯 Functional Correctness | 🟠 Major | ⚡ Quick win

Preset updates will reclassify existing 8-16 GB / 16 GB+ users as "Custom" tier.

current_tier_from_config exact-matches config.{chat,vision,embedding}_model_id against the current preset definitions. Users who already applied the 8-16 GB tier (nomic-embed-text:latest) or the 16 GB+ tier (gemma3n:e4b-it-q8_0) before this change have those old values persisted in config. After this update, neither the selected_tier fast-path nor the all_presets() fallback loop will match, so they'll silently fall through to ModelTier::Custom even though model resolution itself still works fine via the chat allowlist back-compat entry. This is purely a tier-detection/UI regression, but it will confuse users who see their preset badge flip to "Custom" after an update they didn't initiate.

Consider having current_tier_from_config (or a migration step in config load) also recognize the previous generation of model ids for these two tiers, so the tier label survives the model swap.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@src/openhuman/inference/presets.rs` around lines 160 - 211, Update
current_tier_from_config to recognize the previous model identifiers for the
8–16 GB and 16 GB+ presets, including nomic-embed-text:latest and
gemma3n:e4b-it-q8_0, in addition to current preset values. Ensure users with
those persisted configurations retain the corresponding preset tier instead of
falling through all_presets() to ModelTier::Custom.

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@gitbooks/features/model-routing/local-and-byok-models.md`:
- Around line 69-75: Update the tier table’s Embeddings entries to stop
recommending all-minilm for Memory Tree-compatible presets, replacing the 1 GB
and 4–8 GB entries with bge-m3 as used by the other compatible tiers.
- Around line 25-28: Update the comparison table’s “Data leaves your machine”
row to explicitly refer to inference data, such as “Inference data leaves your
machine,” while preserving the existing local, BYOK, and hosted values. Keep the
surrounding explanation of backend-dependent features unchanged.

In `@src/openhuman/inference/local/service/vision_embed.rs`:
- Around line 55-99: Add tests for the vision request flow covering both
`resolve_vision_model_id` failure and `ensure_ollama_model_available` failure,
in addition to the existing disabled-vision coverage. Assert each branch returns
the expected error and updates `vision_state` to `"missing"`, while preserving
the successful path behavior.

---

Outside diff comments:
In `@src/openhuman/inference/presets.rs`:
- Around line 160-211: Update current_tier_from_config to recognize the previous
model identifiers for the 8–16 GB and 16 GB+ presets, including
nomic-embed-text:latest and gemma3n:e4b-it-q8_0, in addition to current preset
values. Ensure users with those persisted configurations retain the
corresponding preset tier instead of falling through all_presets() to
ModelTier::Custom.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro

Run ID: 36892b41-f3f1-4410-824f-9eba2d6abb89

📥 Commits

Reviewing files that changed from the base of the PR and between 9750b80 and 131c7f9.

📒 Files selected for processing (10)

• README.md
• gitbooks/SUMMARY.md
• gitbooks/features/model-routing/README.md
• gitbooks/features/model-routing/local-ai.md
• gitbooks/features/model-routing/local-and-byok-models.md
• src/openhuman/inference/local/service/vision_embed.rs
• src/openhuman/inference/mod.rs
• src/openhuman/inference/model_ids.rs
• src/openhuman/inference/presets.rs
• src/openhuman/inference/vision_models.rs

[M3gA-Mind]

@coderabbitai review

[coderabbitai]

✅ Action performed

Review finished.

Note: CodeRabbit is an incremental review system and does not re-review already reviewed commits. This command is applicable only when automatic reviews are paused.

[M3gA-Mind]

@coderabbitai review

[coderabbitai]

✅ Action performed

Review finished.

Note: CodeRabbit is an incremental review system and does not re-review already reviewed commits. This command is applicable only when automatic reviews are paused.

[coderabbitai]

Actionable comments posted: 2

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@src/openhuman/inference/local/install_piper.rs`:
- Around line 661-662: Update the test setup around the stub-writing flow
containing wipe_shared_install_dir(&config) to create a Drop-based cleanup guard
before any shared-directory mutation, ensuring cleanup runs during unwinding as
well as normal completion. Make the guard perform best-effort cleanup and remove
the existing tail-only wipe_shared_install_dir call.
- Around line 628-643: Update the test’s path resolution so
workspace_piper_binary_candidates and wipe_shared_install_dir use a temporary
shared root under the TempDir rather than the real ~/.openhuman/bin/piper,
regardless of ambient environment variables. Adjust the relevant test setup and
cleanup while preserving shared_install_lock synchronization and existing
assertions.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro

Run ID: 6753a92a-ad96-4a46-ad39-95a669feb09b

📥 Commits

Reviewing files that changed from the base of the PR and between 131c7f9 and 026da83.

📒 Files selected for processing (3)

• gitbooks/features/model-routing/local-and-byok-models.md
• src/openhuman/inference/local/install_piper.rs
• src/openhuman/inference/local/service/vision_embed.rs

🚧 Files skipped from review as they are similar to previous changes (1)

• gitbooks/features/model-routing/local-and-byok-models.md

[M3gA-Mind]

@coderabbitai review

[coderabbitai]

✅ Action performed

Review finished.

Note: CodeRabbit is an incremental review system and does not re-review already reviewed commits. This command is applicable only when automatic reviews are paused.

[M3gA-Mind]

@coderabbitai review

[coderabbitai]

✅ Action performed

Review finished.

Note: CodeRabbit is an incremental review system and does not re-review already reviewed commits. This command is applicable only when automatic reviews are paused.