> ## Documentation Index
> Fetch the complete documentation index at: https://docs.chatzy.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Additional concepts

<Accordion title=" Chunking">
  Chunking is the process of splitting large documents or datasets into smaller, manageable pieces called **chunks**.

  ***

  ### Why Chunking Matters

  AI models can only process a limited number of tokens at once.\
  Chunking ensures that long documents remain **searchable**, **indexable**, and **retrievable**.

  ***

  ### Best Practices

  * **Smaller chunks** → Higher precision (AI focuses better) but may lose broader context
  * **Larger chunks** → Preserve context but may retrieve less relevant information
  * The right balance depends on your data and use case

  ***

  ### Summary

  Chunking makes your data **searchable and structured**, preparing it for retrieval systems like RAG.
</Accordion>

<Accordion title="RAG (Retrieval-Augmented Generation)">
  RAG (Retrieval-Augmented Generation) improves chatbot accuracy by combining **knowledge retrieval** with **LLM generation**.\
  Instead of relying only on the model’s memory, responses are grounded in your uploaded data.

  ***

  ## How RAG Works in Chatzy

  ### 1) Knowledge Base Upload & Processing

  * Users upload their Knowledge Base in the **Data section**
  * The system:
    * Splits content into chunks
    * Generates vector embeddings
    * Stores them in the database

  Chunking can be configured in **Advanced Bot Settings**:

  * Number of chunks
  * Chunk length

  ***

  ### 2) Query Understanding

  When a user sends a message:

  * The system checks if KB(knowledge base) sources are available
  * A lightweight GPT model analyzes:
    * User message
    * Conversation history
  * It generates a **search query** based on user intent

  👉 If no relevant intent is found, the search query may be `null`.

  ***

  ### 3) Retrieval Step

  **If query is null**

  * No KB context is attached
  * The LLM answers normally

  **If query exists**

  * Vector search is performed on the KB
  * Top N relevant chunks are retrieved

  Similarity and relevance thresholds can be configured in **Advanced Settings**.

  ***

  ### 4) Hybrid Search (Optional)

  Hybrid Search combines **similarity search** and **text (keyword) search** into a final relevancy score.

  How it works:

  * Extra chunks are fetched as a broad match
  * This step may include some noise
  * The system then **reranks results**
  * Only the top N most relevant chunks are kept

  This ensures:

  * Better precision
  * Less irrelevant context
  * Higher response quality

  ***

  ### 5) Generation Step

  * Retrieved chunks are combined
  * Added to the system prompt as context
  * Sent to the LLM with instructions

  The LLM then generates a grounded, context-aware response.

  ***

  ## RAG Process Flow

  ```plaintext theme={null}
  User Query
     │
     ▼
  Intent → Search Query Generation
     │
     ▼
  Vector / Hybrid Search on KB
     │
     ▼
  Reranking (Top N Chunks)
     │
     ▼
  Context + Prompt → LLM
     │
     ▼
  Final Answer

  ```

  ***

  ## Benefits of RAG

  ✅ Reduces hallucinations <br />
  ✅ Grounds answers in your data <br />
  ✅ Works well for large knowledge bases <br />
  ✅ Configurable retrieval behavior <br />
  ✅ More accurate and trustworthy responses

  ***

  **In short:**
  Chunking makes your data searchable, and RAG ensures the model uses the *right* information at the *right* time.
</Accordion>

<Accordion title="Advance Settings for Conversational AI Agent">
  ### Location Request

  If your chatbot needs the user’s live location during the conversation, you can instruct it to send a **location request button**. This is useful for scenarios such as booking a service visit, delivery verification, or assigning the nearest agent.

  **Important:**\
  You must clearly mention the **use case inside your Base Prompt**, explaining *when* the bot should send this JSON.

  Use this when:

  * You need the user's address or live location for delivery, service visits, event check-ins, etc.

  **JSON format (use exactly as shown):**

  ```json theme={null}
  {
        "type": "location_request", 
        "content": "Replace with the message that you want to display"
  }
  ```

  ***

  ### Call Permission Request

  Use this when your AI agent needs to **ask the user for permission to place a WhatsApp call** (required for business-initiated calls).

  Why it's needed:

  * WhatsApp requires explicit user consent before initiating an outbound call.

  **Best used for:**

  * Sales follow-ups
  * Demo or consultation calls
  * Appointment confirmation calls

  **Important:**\
  Specify **when the bot should request permission** inside your Base Prompt along with json so it triggers appropriately.(e.g., “If the user requests a callback, use ONLY the following for call permission request”).

  **JSON format:**

  ```json theme={null}
  {
        "type": "call_permission_request", 
        "content": "Replace with the message that you want to display"
  }
  ```

  ***

  ### WhatsApp Call Button

  This option allows the agent to display a **Call Now button** directly inside the conversation.

  Use this when:

  * You want the *user* to start the call themselves (no approval template required).
  * Ideal for support lines, quick escalation, or urgent help.

  When the user taps the button, the call is placed immediately from their device.

  **Important:**\
  In your Base Prompt, describe **in which situations** this button should appear (e.g., “If the user asks to talk to support, use ONLY the following for call button”).

  **JSON format:**

  ```json theme={null}
  {
        "type": "voice_call", 
        "content": "Replace with the message that you want to display"
  }
  ```
</Accordion>