Skip to content

minimax-quirks.md

Latest commit

 

History

History
67 lines (40 loc) · 2.88 KB

minimax-quirks.md

File metadata and controls

67 lines (40 loc) · 2.88 KB
Raw

MiniMax M2.5 API Quirks

MiniMax M2.5 is used as the sole LLM provider via an OpenAI-compatible API. This document tracks compatibility issues and workarounds.

Endpoint

Base URL: https://api.minimaxi.chat/v1

Note the extra "i" — minimaxi not minimax. Using the wrong URL gives silent failures or DNS errors.

Pricing

All standard models cost the same: $0.30/M input tokens, $1.20/M output tokens. No cheaper model exists.

Known Issues

Single System Message Only

MiniMax returns error 2013 ("invalid chat setting") if a request contains more than one system message.

Workaround: Merge all system-level instructions into a single system message. When adding JSON formatting instructions, append to the existing system message rather than adding a new one.

No response_format Support

Passing response_format: {"type": "json_object"} causes error 2013.

Workaround: Append "You MUST respond with valid JSON only. No markdown fences, no explanation." to the system message. Parse the response and strip markdown code fences if present.

<think> Tags in Responses

MiniMax M2.5 sometimes includes <think>...</think> blocks containing chain-of-thought reasoning in the response content.

Workaround: Strip with regex after receiving response:

import re
content = re.sub(r'<think>[\s\S]*?</think>', '', content).strip()

No Logprobs

MiniMax does not support the logprobs parameter. Any reranking or confidence scoring must use alternative approaches.

Workaround: adapter/minimax_reranker.py uses an LLM-based reranking prompt instead of logprob-based scoring.

Embedding Rate Limits

The embo-01 embedding model has very aggressive RPM limits on the free tier, making it impractical for batch operations like graph building.

Workaround: Use local fastembed with BAAI/bge-small-en-v1.5 (ONNX, runs on CPU). See adapter/local_embedder.py.

Stale Connections

Long-running processes (like OASIS simulation) can experience hung HTTP connections to the MiniMax API. The server closes the connection but the client doesn't detect it, causing the process to block indefinitely.

Workaround: No clean fix yet. Force-kill stuck processes after timeout. The simulation data in SQLite is intact even when the process hangs.

Files Modified for MiniMax Compatibility

File Change
vendor/mirofish/backend/app/utils/llm_client.py Removed response_format, added <think> stripping, JSON instruction merging
vendor/mirofish/backend/app/services/simulation_config_generator.py Removed response_format, added <think> stripping
vendor/mirofish/backend/app/services/oasis_profile_generator.py Removed response_format, added <think> stripping
adapter/minimax_llm_client.py Custom OpenAIClient subclass for Graphiti
adapter/minimax_reranker.py LLM-based reranker (no logprobs)