Pipeline และ Policies ของ Rasa คือหัวใจสำคัญที่ช่วยให้ระบบสามารถประมวลผลข้อความผู้ใช้ แยกแยะเจตนา (intent) และระบุข้อมูลเฉพาะ (entities) รวมถึงกำหนดการตอบสนองที่เหมาะสมได้ ไม่ว่าคุณจะสร้างแชทบอทเพื่อให้บริการลูกค้า ผู้ช่วยเสมือน หรือ AI สนทนา การเข้าใจการทำงานของ Pipeline และ Policies จะช่วยให้คุณออกแบบแชทบอทที่มีประสิทธิภาพและตอบโจทย์การใช้งานได้ดียิ่งขึ้น

Continue reading "การปรับแต่ง Pipeline และ Policies ของ Rasa: คู่มือสำหรับการสร้างแชทบอทที่ชาญฉลาดขึ้น" →

Rasa’s pipeline and policies are at the core of its ability to process user inputs, classify intents, extract entities, and determine the next best action. Whether you’re building a chatbot for customer support, a virtual assistant, or any conversational AI, understanding how these components work will help you design a smarter and more efficient bot.

In this blog post, we’ll break down the pipeline components, explain the role of policies, and include a visual Mermaid.js diagram to show how everything connects.

---

What is a Rasa Pipeline?

The Rasa pipeline is a sequence of components that processes user input and prepares it for intent classification and entity recognition. These components handle tokenization, feature extraction, and more, creating a structured representation of the text.

Think of the pipeline as a conveyor belt, where each component performs a specific task in the text processing workflow.

---

Key Components of the Pipeline

1.Tokenizer

• Breaks user input into smaller units (tokens) like words or characters.
• Critical for languages like Thai, which do not use spaces between words.

Example:

   - name: "custom_components.thai_tokenizer.ThaiTokenizer"
     model: "newmm"

2.Featurizers

• Convert tokens into numerical representations (vectors).
• Example components:
  • CountVectorsFeaturizer: For word or character n-grams.
  • RegexFeaturizer: For pattern-based features like phone numbers or dates.

Example:

   - name: CountVectorsFeaturizer
     analyzer: "char_wb"
     min_ngram: 2
     max_ngram: 4

3.Entity Extractors

• Extract structured data like names, locations, or dates.
• Example components:
  • DucklingEntityExtractor: Automatically detects dates, times, and numbers.
  • RegexEntityExtractor: Captures entities using regex patterns.

Example:

   - name: DucklingEntityExtractor
     dimensions: ["time", "number"]

4.Intent Classifier

• Identifies the intent of the user’s input and extracts entities simultaneously using the DIETClassifier.

Example:

   - name: DIETClassifier
     epochs: 100
     entity_recognition: True

5.Fallback Mechanism

• Handles low-confidence predictions to avoid incorrect responses.

Example:

   - name: FallbackClassifier
     threshold: 0.3

---

Policies: Controlling Dialogue Flow

While the pipeline processes user inputs, policies determine the bot's next action. They decide whether the bot should follow a rule, recall a predefined path, or generalize based on context.

Common Policies in Rasa

1.RulePolicy

• Handles predictable flows and FAQs.

Example:

   - name: RulePolicy
     core_fallback_threshold: 0.4
     enable_fallback_prediction: True

2.MemoizationPolicy

• Remembers exact conversation paths from training stories.

3.TEDPolicy

• Generalizes to predict the next action when the conversation deviates from training stories.

Example:

   - name: TEDPolicy
     max_history: 5
     epochs: 100

4.FallbackPolicy

• Triggers a fallback action when confidence is too low.

---

How It All Works: A Visual Representation

Below is a Mermaid.js diagram showing how the pipeline and policies interact to process user inputs and generate responses:

graph TD
    A[User Input] -->|Raw Text| B[Tokenizer]
    B -->|Tokens| C[Featurizers]
    C -->|Features| D[Entity Extractors]
    C -->|Features| E[Intent Classifier]
    D -->|Entities| F[DIETClassifier]
    E -->|Intent| F[DIETClassifier]
    F -->|Predictions| G[Policy Decision]

    G -->|Follows Rules| H[RulePolicy]
    G -->|Known Paths| I[MemoizationPolicy]
    G -->|Generalized| J[TEDPolicy]
    G -->|Fallback| K[FallbackPolicy]

    H --> L[Bot Action]
    I --> L
    J --> L
    K --> L
    L --> M[Bot Response]

    %% Additional Notes
    subgraph Rasa Pipeline
        B
        C
        D
        E
        F
    end

    subgraph Rasa Policies
        H
        I
        J
        K
    end

---

Example: Building a Pipeline for Thai

Here’s an example pipeline tailored for the Thai language, which has unique tokenization and feature extraction requirements:

language: th

pipeline:
  - name: "custom_components.thai_tokenizer.ThaiTokenizer"
    model: "newmm"
  - name: RegexFeaturizer
  - name: CountVectorsFeaturizer
    analyzer: "char_wb"
    min_ngram: 2
    max_ngram: 4
  - name: DucklingEntityExtractor
    dimensions: ["time", "number", "amount-of-money"]
  - name: DIETClassifier
    epochs: 100
    entity_recognition: True
  - name: FallbackClassifier
    threshold: 0.3

---

Tips for Optimization

1.Start Simple:

• Begin with essential components (e.g., tokenizer, featurizers, DIETClassifier).
• Add advanced features like LanguageModelFeaturizer or custom components later.

2.Validate Data:

• Use rasa data validate to catch inconsistencies in your training data.

3.Monitor Performance:

• Use rasa test to evaluate the bot's performance and refine as needed.

---

Conclusion

Mastering Rasa’s pipeline and policies allows you to build a chatbot that processes user inputs efficiently and responds intelligently. By combining well-optimized pipelines with clear dialogue policies, you can create a bot that’s accurate, flexible, and tailored to your use case.

Whether you’re building for Thai or any other language, start simple, test iteratively, and refine your configurations to achieve the best results.

Let us know if you have any questions or need help with your pipeline! 😊

---

Feel free to share feedback or ask for more detailed examples.