lang-agent/README.md

Lang Agent Chat API

这是一个基于FastAPI的聊天API服务，使用OpenAI格式的请求来调用pipeline.invoke方法进行聊天。

安装依赖

# recommended to install as dev to easily modify the configs in ./config
python -m pip install -e .

环境变量

make a .env with:

ALI_API_KEY=<ALI API KEY>
ALI_BASE_URL="https://dashscope.aliyuncs.com/compatible-mode/v1"
MCP_ENDPOINT=<OBSOLETE>
LANGSMITH_API_KEY=<LANG SMITH API KEY> # for testing only; location of where dataset is stored

Hardware tools

update the link to xiaozhi server in configs/mcp_config.json

Configure for Xiaozhi

0. Start the fastapi_server/server_dashscope.py file
1. Make a new model entry in xiaozhi with AliBL as provider.
2. Fill in the base_url entry. The other entries (API_KEY, APP_ID) can be garbage
   • for local computer base_url=http://127.0.0.1:8588/api/
   • if inside docker, it needs to be base_url=http://{computer_ip}:8588/api/

运行服务

API key setup

server_dashcop.py and server_openai.py both require api key; generate one and set

FAST_AUTH_KEYS=API_KEY1,API_KEY2    # at least one

FAST_AUTH_KEYS will be used as the api-key for authentication when the api is requested.

# for easy debug; streams full message internally for visibility
python fastapi_server/fake_stream_server_dashscopy.py

# for live production; this is streaming
python fastapi_server/server_dashscope.py

# start server with chatty tool node; NOTE: streaming only!
python fastapi_server/server_dashscope.py route chatty_tool_node

# this supports openai-api; 
python fastapi_server/server_openai.py

see sample usage in fastapi_server/test_dashscope_client.py to see how to communicate with fake_stream_server_dashscopy.py or server_dashscope.py service

Conversation Viewer

A web UI to visualize and browse conversations stored in the PostgreSQL database.

Setup

1. Ensure your database is set up (see scripts/init_user.sql and scripts/recreate_table.sql)
2. Set the CONN_STR environment variable:

   export CONN_STR="postgresql://myapp_user:secure_password_123@localhost/ai_conversations"
   

Running the Viewer

python fastapi_server/server_viewer.py

Then open your browser and navigate to:

http://localhost:8590

Features

• Left Sidebar: Lists all conversations with message counts and last updated timestamps
• Main View: Displays messages in a chat-style interface
  • Human messages appear on the right (blue bubbles)
  • AI messages appear on the left (green bubbles)
  • Tool messages appear on the left (orange bubbles with border)

The viewer automatically loads all conversations from the messages table and allows you to browse through them interactively.

Openai API differences

For the python openai package it does not handle memory. Ours does, so each call remembers what happens previously. For managing memory, pass in a thread_id to manager the conversations

from openai import OpenAI

client = OpenAI(
        base_url=BASE_URL,
        api_key="test-key"  # see put a key in .env and put it here; see above
    )

client.chat.completions.create(
            model="qwen-plus", 
            messages=messages,
            stream=True,
            extra_body={"thread_id":"2000"}  # pass in a thread id; must be string
        )

Runnables

everything in scripts:

• For sample usage see scripts/demo_chat.py.
• To evaluate the current default config scripts/eval.py
• To make a dataset for eval scripts/make_eval_dataset.py

Registering MCP service

put the links in configs/mcp_config.json

Graph structure

Graph structure:

We choose this structure to overcome a limitation in xiaozhi. Specifically, both normal chatting and tool use prompts are deligated to one model. That leads to degregation in quality of generated conversation and tool use. By splitting into two model, we effectively increase the prompt limit size while preserving model quality.

Modifying LLM prompts

Refer to model above when modifying the prompts.
they are in configs/route_sys_prompts

• chat_prompt.txt: controls chat_model_call
• route_prompt.txt: controls router_call
• tool_prompt.txt: controls tool_model_call
• chatty_prompt.txt: controls how the model say random things when tool use is in progress. Ignore this for now as model architecture is not yet configurable

Stress Test results

Dashscope server summary

Non-Streaming

Concurrency	Requests	Success %	Throughput (req/s)	Avg Latency (ms)	p95 (ms)	p99 (ms)
1	10	100.00%	0.77	1293.14	1460.48	1476.77
5	25	100.00%	2.74	1369.23	1827.11	3336.25
10	50	100.00%	6.72	1344.48	1964.75	2165.77
20	100	100.00%	10.90	1688.06	2226.49	2747.19
50	200	100.00%	11.75	3877.01	4855.45	5178.52

Streaming

Concurrency	Requests	Success %	Throughput (req/s)	Avg Latency (ms)	p95 (ms)	p99 (ms)
1	10	100.00%	0.73	1374.08	1714.61	1715.82
10	50	100.00%	5.97	1560.63	1925.01	2084.21
20	100	100.00%	9.28	2012.03	2649.72	2934.84

Interpretation - Handling concurrently 20 conversations should be ok