cloud/music_library
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
Files
e36d80818e658e483b67d0c1663fc8bf1581aeff
music_library/backlog/docs/doc-22 - Evaluation-Replacing-OpenAI-with-Mistral.md
T
Claudio Ortolina 544bf44ec6 Research swapping OpenAI with Mistral
2026-05-14 11:37:12 +01:00

12 KiB
Raw Blame History

id, title, type, created_date, updated_date, tags
id title type created_date updated_date tags
doc-22 Evaluation: Replacing OpenAI with Mistral guide 2026-05-14 10:32 2026-05-14 10:35
openai
mistral
evaluation
migration
api-integration

Evaluation: Replacing OpenAI with Mistral

Summary

Overall complexity: MEDIUM-HIGH (unchanged, but web search picture clarified)

The discovery that Mistral offers web search via their Studio Agents API (a separate product from Chat Completions / La Plateforme) changes the web search calculus. It does NOT eliminate the gap — it shifts it to an architectural decision between two harder paths.

OpenAI Integration Scope

5 call sites across 3 core capabilities:

Capability Module Function OpenAI API endpoint
Genre population (structured JSON) Records.Enrichment populate_genres/1 /v1/chat/completions
Record similarity (embeddings) Records.Similarity generate_embedding/1 /v1/embeddings
Chat streaming (SSE + web search) Chats.RecordChat stream_response/3 /v1/responses
Chat streaming (SSE + web search) Chats.ArtistChat stream_response/3 /v1/responses
Chat streaming (SSE + web search) Chats.CollectionChat stream_response/3 /v1/responses

Supporting modules (7 files): OpenAI facade, OpenAI.API, OpenAI.Config, OpenAI.Completion, OpenAI.API.ErrorResponse, Worker.ErrorHandler, RetryDelay.

Tests: 344 lines in dedicated OpenAI tests, plus references in 8 other test files.

Config: 3 files (config.exs, runtime.exs, test.exs).

Complexity by Capability

1. Chat Streaming — HIGH 🔴

The hardest part. Current code uses OpenAI's Responses API (/v1/responses), not Chat Completions. Mistral has two separate APIs, each with different trade-offs:

Option A: Mistral Chat Completions API (La Plateforme)

Use /v1/chat/completions with stream: true. SSE format is different from OpenAI's Responses API — emits choices[0].delta.content instead of response.output_text.delta. The decode_responses_event/2 function needs a full rewrite.

BUT: No web search support. Mistral's Chat Completions API does NOT have a web search tool. To keep web search, you'd need to pre-fetch via BraveSearch and inject results as context — a material feature change.

Option B: Mistral Studio Agents API

Use the Studio API (client.beta.agents.create + client.beta.conversations.start). This is a completely different API surface from La Plateforme:

  • Requires agent creation before starting conversations
  • Uses tools: [{"type": "web_search"}] or "web_search_premium"
  • Different SSE event format (tool.execution + message.output interleaved chunks)
  • Different base URL, different auth, possibly different rate limiting
  • Includes citations (references with URLs/titles) — a nice upgrade
  • Streaming plumbing (into: :self) is reusable, but the event decoder is entirely different

This is a materially larger rewrite than Option A, but preserves web search natively.

SSE Event Format Comparison

Event type OpenAI Responses API Mistral Chat Completions Mistral Studio Agents
Text delta response.output_text.delta choices[0].delta.content Chunks within message.output entries
Error error / response.failed Standard HTTP error Tool execution errors
Tool execution N/A (inline in response) N/A tool.execution entries
Citations N/A N/A tool_reference chunks with URLs

Model Differences

CollectionChat uses gpt-5.1 (very large context). Mistral equivalent for large context would be mistral-large-latest. For the Studio Agents API, the docs show mistral-medium-latest used with web search.

2. Embeddings — MEDIUM 🟡

Both use /v1/embeddings with near-identical response shapes. Mistral uses mistral-embed vs OpenAI's text-embedding-3-small.

Key concern: Dimension mismatch. OpenAI's text-embedding-3-small produces 1536-dimensional vectors. Mistral's mistral-embed produces 1024-dimensional vectors. The record_embeddings table stores these via sqlite-vec. This means:

  • All existing embeddings must be regenerated (mass re-embed)
  • vec_f32() and vec_distance_cosine work with any dimension — no schema change needed
  • Just trigger RecordGenerateAllEmbeddings after migration

3. Genre Population — LOW 🟢

Both use Chat Completions with response_format: %{type: "json_object"}. The OpenAI.Completion struct maps cleanly. Response shape (choices[0].message.content) is identical. Nearly a drop-in replacement — just change model name and base URL.

4. Error Handling — MEDIUM 🟡

If using Studio Agents API, error handling becomes more complex due to the multi-step nature (agent creation can fail separately from conversation streaming). Error body shapes differ between La Plateforme and Studio APIs.

5. Configuration & Tests — MEDIUM 🟡

If using both La Plateforme (for embeddings + genres) AND Studio API (for chat), you're now managing two Mistral API keys, two base URLs, and two sets of rate limits. Tests need stubs for both API surfaces.

Mistral API Surface Map

Need OpenAI solution Mistral La Plateforme Mistral Studio Agents
Chat completions (JSON) /v1/chat/completions ✅ /v1/chat/completions ❌ Different API
Chat streaming /v1/responses (SSE) ✅ /v1/chat/completions (SSE) ✅ Conversations API (SSE)
Web search web_search_preview tool ❌ Not available ✅ web_search / web_search_premium
Embeddings /v1/embeddings ✅ /v1/embeddings ❌ Not available
Citations in response ❌ Not built-in ❌ Not built-in ✅ tool_reference chunks

Revised Decision Matrix

Path Web search Effort Risk Note
A: La Plateforme only (Chat + Embeddings) ❌ Lost or manual BraveSearch injection ~16 hours Medium Simpler code, but chat loses native web search
B: Hybrid (La Plateforme for embeddings/genres, Studio for chat) ✅ Native ~24+ hours High Two API keys, two API surfaces, complex error handling
C: Studio only (all via Agents API) ✅ Native ~28+ hours High Embeddings still need La Plateforme; Studio can't do embeddings

Recommended path: Option A with BraveSearch injection for web search. This keeps the architecture simple (one API surface), preserves web search functionality by pre-fetching search results via the already-integrated BraveSearch API, and avoids the complexity of managing two Mistral API products.

Files Affected

New files to create

  • lib/mistral.ex — Facade (replaces OpenAI)
  • lib/mistral/api.ex — HTTP client (La Plateforme)
  • lib/mistral/api/config.ex — NimbleOptions config schema
  • lib/mistral/api/error_response.ex — implements MusicLibrary.ErrorResponse
  • lib/mistral/completion.ex — Struct for non-streaming completions
  • test/mistral_test.exs — Facade tests
  • test/mistral/api_test.exs — API tests
  • test/support/fixtures/mistral_fixtures.ex — Response fixtures

Files to modify

  • lib/music_library/chats/record_chat.ex — OpenAI.chat_stream → Mistral.chat_stream; potentially add BraveSearch injection
  • lib/music_library/chats/artist_chat.ex — same
  • lib/music_library/chats/collection_chat.ex — same
  • lib/music_library/records/similarity.ex — OpenAI.embeddings → Mistral.embeddings
  • lib/music_library/records/enrichment.ex — OpenAI.gpt → Mistral.gpt
  • lib/music_library/retry_delay.ex — Add Mistral reset header parsing
  • lib/music_library/worker/error_handler.ex — Add Mistral.API.ErrorResponse to @error_structs
  • config/config.exs — Replace OpenAI config with Mistral
  • config/runtime.exs — Read MISTRAL_API_KEY env var
  • config/test.exs — Stub Mistral.API via Req.Test

Files to remove

  • lib/open_ai.ex
  • lib/open_ai/api.ex
  • lib/open_ai/api/config.ex
  • lib/open_ai/api/error_response.ex
  • lib/open_ai/completion.ex
  • test/open_ai_test.exs
  • test/open_ai/api_test.exs

Migration Risk Matrix

Risk Severity Mitigation
Web search feature gap (if using La Plateforme) High Pre-fetch via BraveSearch, inject as context; or use Studio Agents API
Two-API-surface complexity (if using Studio) High Start with La Plateforme only; add Studio later if needed
Chat quality regression Medium Compare responses side-by-side; keep OpenAI key active during transition
Embedding dimension mismatch Medium Full re-embed via existing batch worker; verify similarity results
Rate limit header format Low Test with actual Mistral 429 responses; adjust header parsing
Model prompt compatibility Medium System prompts may need tuning for Mistral models

Estimated Effort (Option A: La Plateforme with BraveSearch injection)

Task Hours
Create Mistral module tree (facade, API, config, error) 3-4
Rewrite SSE event decoder for Chat Completions format 2-3
Web search via BraveSearch injection into chat context 3-4
Update 5 call sites (chat × 3, embeddings, genres) 1-2
Error handling + RetryDelay updates 1
Config changes (3 files) 0.5
Tests: new + update existing 3-4
Regenerate all embeddings (automated, no manual work)
Integration testing (chat quality, similarity) 2-3
Documentation update (architecture doc) 0.5
Total ~16-24 hours

References

Reference in New Issue View Git Blame Copy Permalink