Skip to main content

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.

The Advanced Settings tab gives you detailed control over how your chatbot processes information. These options affect how data is retrieved, how conversations are remembered, and how responses are generated.

​
Key Features

  • Adjust Chunks
    Control how your uploaded data is split into smaller units (chunks) for the AI to process.
    • Chunks: Number of segments your data is divided into (e.g., 5).
    • Chunk Length: Size of each chunk (measured in characters).
    • Formula:
      # Chunks × Chunk Length = Context Length
      Example: 3 × 1500 = 4500 chars
    • Tip: Smaller chunks improve precision but may reduce contextual understanding. Larger chunks preserve more information but can introduce irrelevant details.
  • Hybrid Search
    Improves relevancy in knowledge search by combining keyword and semantic matching.
    • How it Works:
      • Retrieves a larger set of chunks:
        # Chunks × Initial Fetch Multiplier (default: 5).
      • Runs a text search across these expanded results.
      • Combines similarity search + text search into a final relevancy score.
    • Alpha Parameter (0–1):
      • Alpha = 1 → similarity-focused (semantic emphasis).
      • Alpha = 0 → text-focused (keyword emphasis).
      • Values in between balance both methods (default: 0.5).
    • Use Case:
      Adjust for your needs — use higher alpha for conversational queries, lower alpha for exact keyword lookups.
    Customize Hybrid Search to fine-tune how your bot balances meaning vs. exact match.
  • Add Historical Context for KB Query Construction
    Turn this option on when you want to construct a knowledge base search query specific to exact key terms or entities from the user’s message.
    • Benefit: Improves continuity, relevance, and personalized responses.
    • Initial Fetch Multiplier: Defines how many more documents to fetch in the initial retrieval stage before ranking (default: 5).
    • Alpha: Balances the weight between query relevance and conversation context when scoring retrieved results (default: 0.5).
    • Credits: Consumes additional credits (0.2 credit per message).
If this option is off, the user’s query is used directly to search the knowledge base, which might result in less relevant chunks being picked.
When you turn this on, we use an LLM along with historical context to construct the query. This results in more relevant chunks being selected from the knowledge base, ultimately producing more coherent and context-aware chatbot replies.

​
Prompt Example: Using Historical Context for Query Construction

### Example: Query Construction Using Historical Context  

Instruction:
Analyze the recent conversation history to identify the user’s current intent and construct an optimal knowledge base retrieval query. Focus on the most recent exchanges to determine the immediate need.  

Query Construction Requirements:
- Extract key terms, keywords, and specific entities from user messages.  
- Include relevant technical terms, process names, and contextual details.  
- For requests involving documents, payments, or mail drafts → ensure entity names (e.g., company, product, institution) are included.  
- Construct queries that maximize the chances of retrieving targeted knowledge base content.  
- If the user intent is unclear or unrelated to KB topics, return `"null"`.  

Output: 
Provide the constructed query with essential keywords and terms needed for accurate knowledge base retrieval.
  • Max Tokens
    Sets the maximum length of a response (measured in tokens – words or word parts).
    Helps manage cost, control verbosity, and ensure the bot responds within limits.
  • Token Distribution
    Visual slider that shows how your total Max Tokens are allocated across different components:
    • Base Prompt + Function Call + Media
    • Context (Knowledge)
    • Chat History
    • Response + Reasoning
    Example allocation:
    • Base Prompt + Function Call + Media → ~1,500 tokens
    • Context (Knowledge) → ~7,500 tokens
    • Chat History → ~5,000 tokens
    • Response + Reasoning → ~1,000 tokens
    You can dynamically adjust these allocations to balance accuracy, context retention, and reasoning depth.