From 8b07e76d67b4f31e0cdf944ebff0958314945cba Mon Sep 17 00:00:00 2001 From: cnphpbb Date: Fri, 19 Jun 2026 21:21:46 +0800 Subject: [PATCH] fix(honcho): override deduction/induction specialist model to deepseek/siliconflow MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit honcho-deriver dreamer 模块的 deduction/induction specialist 有独立的 model config（src/config.py:1189-1208），默认值 hardcode gpt-5.4-mini + openai SDK 默认 endpoint（api.openai.com/v1）。原 stack.yml 只覆盖了 DERIVER_MODEL_CONFIG 和 EMBEDDING_MODEL_CONFIG，导致 dream cycle 在 国内持续 timeout。 - stack.yml deriver 段显式注入 DEDUCTION_MODEL_CONFIG__* / INDUCTION_MODEL_CONFIG__*（model=deepseek-ai/DeepSeek-V4-Flash， base_url=https://api.siliconflow.cn/v1） - DERIVER_MODEL_CONFIG__OVERRIDES__BASE_URL 同步加上 - env.cfg.example: 同步 HONCHO_DEDUCTION_MODEL / HONCHO_INDUCTION_MODEL 变量；HONCHO_DERIVER_MODEL 默认值从 gpt-5.4-mini 改为 deepseek-ai/DeepSeek-V4-Flash；HONCHO_LLM_BASE_URL 默认从 openai.com 改为 siliconflow.cn 诊断依据：vm61 docker inspect honcho-deriver + logs 显示 src.dreamer.orchestrator 仍在请求 openai/gpt-5.4-mini。 .env.* --- honcho/env.cfg.example | 9 +++++++-- honcho/stack.yml | 9 ++++++++- 2 files changed, 15 insertions(+), 3 deletions(-) diff --git a/honcho/env.cfg.example b/honcho/env.cfg.example index e17cf0a..7a6c8fb 100644 --- a/honcho/env.cfg.example +++ b/honcho/env.cfg.example @@ -61,14 +61,19 @@ HONCHO_EMBEDDING_BASE_URL=https://api.siliconflow.cn/v1 # 必须与已有 DB schema 的 public.documents.embedding dim 一致；切换模型必须先 migrate 旧数据 # 若 Honcho 自动推断失败，需在 stack.yml 同步设置 EMBEDDING_VECTOR_DIMENSIONS=1536 # deriver 必须用支持 tool calling 的模型 -HONCHO_DERIVER_MODEL=gpt-5.4-mini +HONCHO_DERIVER_MODEL=deepseek-ai/DeepSeek-V4-Flash +# dreamer 的 deduction/induction specialist（默认走 OpenAI SDK + gpt-5.4-mini， +# 会请求 api.openai.com/v1 在国内超时）；必须显式覆盖到国内可达端点 +HONCHO_DEDUCTION_MODEL=deepseek-ai/DeepSeek-V4-Flash +HONCHO_INDUCTION_MODEL=deepseek-ai/DeepSeek-V4-Flash HONCHO_DERIVER_WORKERS=1 # ============================================================ # LLM（OpenAI 兼容协议；可换 OpenRouter / vLLM / Ollama 等端点） # deriver 启动时强校验 API Key 非空，占位符会反复重启 +# 注意：必须同时被 deriver/dreamer/embedding 复用，否则 specialist 会回落 openai.com # ============================================================ -HONCHO_LLM_BASE_URL=https://api.openai.com/v1 +HONCHO_LLM_BASE_URL=https://api.siliconflow.cn/v1 # ============================================================ # 敏感值（必填，env.cfg 不提交）— 在 .gitignore 已忽略 diff --git a/honcho/stack.yml b/honcho/stack.yml index bada9fc..486ffb4 100644 --- a/honcho/stack.yml +++ b/honcho/stack.yml @@ -93,7 +93,14 @@ services: - CACHE_ENABLED=true - NAMESPACE=${HONCHO_NAMESPACE:-honcho} - DERIVER_MODEL_CONFIG__TRANSPORT=openai - - DERIVER_MODEL_CONFIG__MODEL=${HONCHO_DERIVER_MODEL:-gpt-5.4-mini} + - DERIVER_MODEL_CONFIG__MODEL=${HONCHO_DERIVER_MODEL:-deepseek-ai/DeepSeek-V4-Flash} + - DERIVER_MODEL_CONFIG__OVERRIDES__BASE_URL=${HONCHO_LLM_BASE_URL:-https://api.siliconflow.cn/v1} + - DEDUCTION_MODEL_CONFIG__TRANSPORT=openai + - DEDUCTION_MODEL_CONFIG__MODEL=${HONCHO_DEDUCTION_MODEL:-deepseek-ai/DeepSeek-V4-Flash} + - DEDUCTION_MODEL_CONFIG__OVERRIDES__BASE_URL=${HONCHO_LLM_BASE_URL:-https://api.siliconflow.cn/v1} + - INDUCTION_MODEL_CONFIG__TRANSPORT=openai + - INDUCTION_MODEL_CONFIG__MODEL=${HONCHO_INDUCTION_MODEL:-deepseek-ai/DeepSeek-V4-Flash} + - INDUCTION_MODEL_CONFIG__OVERRIDES__BASE_URL=${HONCHO_LLM_BASE_URL:-https://api.siliconflow.cn/v1} - DERIVER_WORKERS=${HONCHO_DERIVER_WORKERS:-1} - EMBEDDING_MODEL_CONFIG__TRANSPORT=openai - EMBEDDING_MODEL_CONFIG__MODEL=${HONCHO_EMBEDDING_MODEL:-Qwen/Qwen3-Embedding-8B}