Skip to content

Latest commit

 

History

History
101 lines (86 loc) · 3.2 KB

README.md

File metadata and controls

101 lines (86 loc) · 3.2 KB
Raw

Agents

Implementing a Custom Agent

The Agent Class

After collecting dialog data, you may want to build a virtual agent that can automatically handle users' request and reply to them. We provide an Agent as a base for your agent. Customized agents should implement the on_message method. The simplest example is the echo agent, which just repeats the user's message. You can use the chat interface to interact with an agent, and use the agent interface to see what API calls it makes.

The on_message method in the Agent class is called whenever the system receives user-message and returns an agent-message in response.

An example:

In the following example, the user wants to go to a roof top bar in Marina del Rey, CA.

The parameter events is passed to on_message and contains the current dialog context (variables are omitted to save space). See the example in dataset/examples/log_60.json for the complete log.

{
    "events": [
        {
          "event_type": "user_utterance",
          "utterance": "i want to go to a rooftop bar",
          "variables": [...]
        },
        {
          "event_type": "api_call",
          "endpoint": "find_place",
          "params": [
            {
              "param": "query",
              "variable_name": "u1_6 + \" \" + u1_7",
              "value": "rooftop bar"
            },
            {
              "param": "src latitude",
              "variable_name": "source_latitude",
              "value": 33.9816425
            },
            {
              "param": "src longitude",
              "variable_name": "source_longitude",
              "value": -118.4409761
            }
          ],
          "variables": [...]
        }
    ]
}

on_message method returns next agent action(s) as a list of JSON objects

[
    {
        "body": "How about High Rooftop Lounge?",
        "template": "How about {}?",
        "variables": [
            "v1_name"
        ]
    }
]

Agent Training and Evaluation Example

An example detailing the training and evaluation procedure for an agent pre-trained off GPT2 is provideded here. Note that this is how we produced the pre-trained GPT2 agent listed below.

Deploying your agent

Follow the deployment instructions in 30-second-startup, with --agent_class_name and --agent_model_file set to your agent.

Pre-Trained Agents

There is a pre-trained model at . They work for --domain=destination and can be loaded in the backend like so:

cd ../gui/backend
python main.py --port=<backend_port> --domain=destination --num_rooms=1 --log_dir=<your_log_dir> \
    --agent_class_name=$AGENT_CLASS_NAME \
    --agent_model_path=$AGENT_MODEL_PATH

GPT Agent

AGENT_CLASS_NAME=agents.gpt2_agent.agent.GPT2Agent
AGENT_MODEL_PATH=/home/shared/public/meep/gpt_agent

BERT Agent

AGENT_CLASS_NAME=agents.bert_agent.methods.baseline.agent.BERTAgent
AGENT_MODEL_PATH=/home/shared/public/meep/bert_agent

Rule-based Agent

AGENT_CLASS_NAME=agents.rule_based_agent.RuleBasedAgent
AGENT_MODEL_PATH=""