Skip to content

configuration.md

Latest commit

 

History

History
121 lines (89 loc) · 3.1 KB

configuration.md

File metadata and controls

121 lines (89 loc) · 3.1 KB

⚙️ Configuration

setting.txt

Copy the example file and fill in your credentials:

cp setting.txt.example setting.txt

setting.txt is listed in .gitignore and will never be committed to git.

# Your LLM API key
API_KEY=your_api_key_here

# API base URL (OpenAI-compatible endpoint)
BASE_URL=https://api.openai.com/v1

Supported Providers

TabClaw uses the OpenAI-compatible API format (/chat/completions with streaming and tool/function calling). It works with any provider that supports this interface:

Provider BASE_URL Notes
OpenAI https://api.openai.com/v1 gpt-4o, gpt-4-turbo
DeepSeek https://api.deepseek.com/v1 Default: DeepSeek-V3
SiliconFlow https://api.siliconflow.cn/v1 Many open-source models
Ollama (local) http://localhost:11434/v1 Requires tool-call support
Any OpenAI-compatible API your endpoint —

Model Selection

The default model is set in config.py:

DEFAULT_MODEL = "deepseek-ai/DeepSeek-V3"

Change this to any model your provider supports. TabClaw requires a model that supports parallel tool/function calling and streaming. Recommended models:

Model Strengths
deepseek-ai/DeepSeek-V3 Strong reasoning, fast, cost-effective — default
gpt-4o Best instruction following, reliable tool call JSON
deepseek-chat DeepSeek native endpoint alias
qwen2.5-72b-instruct Good open-source alternative via SiliconFlow

DeepSeek V3 Notes

TabClaw includes a specific workaround for DeepSeek V3's tool call behaviour: the model occasionally leaks raw tool-call markup (<｜tool▁call▁begin｜>…) into delta.content during streaming. The executor detects this marker and suppresses the affected chunks, ensuring the chat UI only displays clean reasoning text. If you switch to a different model, this suppression is inert and does not affect output.

Runtime Data

The following files are created automatically at runtime and are gitignored:

File Contents
data/memory.json User memory — preferences, domain knowledge, context, history insights
data/custom_skills.json Custom skills saved from the UI or learned by the skill distiller
uploads/ Temporarily uploaded CSV/Excel files

Memory File Format

{
  "preferences": {
    "output_language": {
      "value": "Chinese",
      "updated": "2026-01-15T10:23:41+00:00"
    }
  },
  "domain_knowledge": {},
  "user_context": {},
  "history_insights": {}
}

Custom Skills File Format

[
  {
    "id": "a1b2c3d4",
    "name": "profit_margin_ranking",
    "description": "Rank products or regions by profit margin (profit / revenue).",
    "prompt": "",
    "code": "# generalised Python code …",
    "parameters": {}
  }
]

Requirements

fastapi
uvicorn[standard]
python-multipart
aiofiles
openai>=1.0
pandas
openpyxl

Install with:

pip install -r requirements.txt

The openai SDK is used as the async HTTP client for any OpenAI-compatible endpoint — no OpenAI account is required.