Dialog Classification

Classifying Sessions

A classification task consists of a topic and a list of classes. The topic is the type of classification you want to perform, and the classes are the possible categories you want to classify the Session into. The topic is only used as a name or label for the classification task and does not affect the classification itself.

You may optionally specify the number of previous messages to consider when classifying the Session (default 4), and whether to persist the classification in the Session’s metadata (default True).

1
classification = await client.memory.classify_session(
2
    session_id,
3
    name="topic",
4
    classes=["class1", "class2", "other"],
5
    last_n=4,
6
    persist=True
7
)

The returned result will be one of the classes you provided, or “other” if none of the classes are a good fit. The classification result is also stored in the Session’s metadata if persist is True.

1
{ "system": { "classes": { "topic": "travel" } } }

And Sessions viewed in the Zep app will be labeled with the classification result.

Adding Custom Instruction

You may optionally provide a custom instruction to the classification task. This instruction will be injected into the Dialog Classification Prompt.

You may want to use this option if you have specific guidelines that you want to communicate to the LLM in addition to our classification prompt.

1
classification = await client.memory.classify_session(
2
   session_id,
3
   name="experience level",
4
   classes=["advanced", "beginner", "dojo"],
5
   instruction="Classify the user's experience level. You may classify experience as 'dojo' only if the user is a black belt."
6
)

Building a Semantic Router with User Intent

Zep’s Session Classifier can be used to build a semantic router that routes user sessions to different agents or chains based on the user’s intent.

1
user: Hello, my phone isn't responding to touch.

1
classification = await client.memory.classify_session(
2
    session_id,
3
    name="intent",
4
    classes=[
5
        "sales interest",
6
        "needs support",
7
        "has payment question",
8
        "other"
9
    ]
10
)
11
12
print(classification)

1
"needs support"

A High-Performance Tool Picker

Using an agent to pick tools can often be slow and inaccurate. Zep’s Session Classifier allows you to pick tools at very low latency and high accuracy. You may then instruct an LLM to use the selected tool and provided Session information.

1
user: What is the capital of France?

1
classification = await client.memory.classify_session(
2
    session_id,
3
    name="tool",
4
    classes=[
5
        "complete math problems using a calculator",
6
        "research topics or find information with a web search",
7
        "no matching tool",
8
    ]
9
)

1
"research topics or find information with a web search"

Classifying Emotions

Zep’s Session Classifier can be used to classify user emotions and store these classifications in the Session’s metadata.

1
AI: We're unfortunately going to reschedule your appointment.
2
user: Is that entirely necessary? I'm very busy.

1
classification = await client.memory.classify_session(
2
    session_id,
3
    name="emotion",
4
    classes=[
5
        "happy",
6
        "sad",
7
        "frustrated",
8
        "angry",
9
        "other"
10
    ]
11
)

1
"frustrated"

Executing Multiple Classifications

Zep supports executing multiple classification tasks against a Session, allowing you to classify a Session into different categories and store these classifications in the Session’s metadata.

1
{
2
  "system": {
3
    "classes": {
4
      "topic": "travel",
5
      "intent": "needs support",
6
      "emotion": "frustrated"
7
    }
8
  }
9
}

Best Practices

1. Use a small number of classes: The more classes you have, the more difficult it is to classify a Session accurately. Recommendation: no more than 5 or 6 classes.
2. Ensure your classes are well separated: If your classes are too similar, the classifier will have a hard time distinguishing between them.
3. Provide a “none”, “other”, or “unknown” class: If none of the classes are a good fit, the classifier should be able to select an option that indicates this.
4. Limit the number of previous messages: The more messages you consider, the longer the classification will take. Additionally, the context may change through the conversation. Recommendation: no more than the prior 4 to 6 messages.
5. Persist the classification: If you want to use the classification result in the future, you should persist the classification in the Session’s metadata. This is the default.