Skip to main content

intro

sidebar_position: 1

Knowledge Base & Document Q&A

Upload documents and enable AI- powered Q&A. Your documents become searchable knowledge that AI can reference when answering customer questions.

What is a Knowledge Base?​

A knowledge base lets you:

  • Upload documents (PDF, DOCX, TXT)
  • Enable AI to search your documents
  • Answer questions using your content
  • Provide accurate, source- backed responses

When customers ask questions, AI searches your documents and provides answers with source citations.

Why Use a Knowledge Base?​

Accurate Answers​

AI answers questions using YOUR content, not generic information.

Source Citations​

Every answer includes which document it came from. Builds trust.

Always Up- to- Date​

Update your documents, and AI automatically uses the latest information.

No Training Required​

Just upload documents - no need to train the AI separately.

How It Works​

Customer asks a question
 
AI searches your documents
 
Finds relevant information
 
Generates answer with sources
 
Customer sees answer + citations

Creating a Knowledge Base​

Step 1: Create Knowledge Base​

  1. Go to Dashboard -> Knowledge Base
  2. Click "Create New Knowledge Base"
  3. Enter a name (e.g., "Product Documentation")
  4. Add a description (optional)

Step 2: Configure AI Settings​

Choose how AI will process your documents:

Knowledge Base Complete View

The Knowledge Base page provides a comprehensive interface for configuring your RAG system. On the left, you'll see configuration sections organized in an accordion layout, and on the right, a live chat preview where you can test your knowledge base. The page also includes a visual workflow builder at the bottom that shows how your RAG pipeline components are connected.

LLM Provider & Model:

  • AI Provider: OpenAI, Cloudflare Workers AI, or DeepSeek
  • AI Model: Choose a model (e.g., GPT-5, GPT-4o, Llama 3.1 8B/70B, DeepSeek Chat/Coder)

Embeddings & Vector Store:

  • Embedding Model: Voyage AI 3 Large for document search
  • Vector Store: ChromaDB Cloud (serverless) - requires API key, Tenant ID, and Database name

Document Processing:

  • Parser: Default (serverless) or Reducto AI for advanced parsing

Response Settings:

  • Temperature: Control creativity (0.0 = factual, 2.0 = creative)
  • Max Tokens: Limit response size (default: 2000)
  • Top-K: Number of document chunks to retrieve (default: 5)
tip

Start with default settings. You can always adjust later!

Step 3: Upload Documents​

  1. Scroll to the "Document Upload" section
  2. Click the upload area or drag and drop files
  3. Select files (PDF, DOCX, TXT, DOC, MD - Max 10MB per file)
  4. Wait for processing (usually 1-2 minutes per document)

Knowledge Base Page

The document upload section is at the bottom of the Knowledge Base page. You can drag and drop files or click to browse.

What happens:

  • Text is extracted from your document
  • Document is split into searchable chunks
  • AI creates searchable embeddings using Voyage AI
  • Document is stored securely in ChromaDB Cloud

Step 4: Connect to Chat Interface​

  1. Select a chat interface to connect
  2. Your knowledge base will power RAG queries
  3. Test it by asking questions in the chat!

Supported File Types​

PDF Files​

  • Standard PDFs
  • Text- based PDFs work best
  • Scanned PDFs may need OCR first

DOCX Files​

  • Microsoft Word documents
  • Preserves formatting
  • Tables and lists supported

TXT Files​

  • Plain text files
  • UTF- 8 encoding
  • Simple and fast
info

Maximum file size: 10MB per file. For larger documents, split them into multiple files.

Using Your Knowledge Base​

In Chat Interface​

Once connected, customers can ask questions like:

  • "What are your return policies?"
  • "How do I reset my password?"
  • "What features are included in the Pro plan?"

AI will:

  1. Search your documents
  2. Find relevant information
  3. Generate an answer
  4. Cite the source document

Example Response​

Customer: "What's your refund policy?"

AI: "Our refund policy allows returns within 30 days of purchase. 
You can request a refund by contacting support@yourcompany.com.

Source: Returns Policy.pdf, page 3"

Best Practices​

Use Clear Documents​

  • Well- structured documents work best
  • Use headings and sections
  • Keep content organized

Update Regularly​

  • Keep documents current
  • Remove outdated information
  • Add new content as needed

Test Questions​

  • Test common questions
  • Verify answers are accurate
  • Improve documents based on results

Organize by Topic​

  • Create separate knowledge bases for different topics
  • Product docs, support docs, policies, etc.
  • Connect relevant knowledge bases to relevant interfaces

Managing Documents​

View Uploaded Documents​

  • See all documents in your knowledge base
  • View processing status
  • Check document metadata

Remove Documents​

  • Remove outdated documents
  • Documents are removed from search immediately
  • No impact on existing conversations

Update Documents​

  • Upload a new version
  • Old version is automatically replaced
  • AI uses the latest version

Understanding RAG (Retrieval Augmented Generation)​

RAG (Retrieval Augmented Generation) is how AI answers questions using your documents:

  1. Retrieval: AI searches your documents for relevant information
  2. Augmentation: AI combines found information with its knowledge
  3. Generation: AI generates an answer using both sources

This means AI answers are:

  • Based on YOUR content
  • Accurate and up- to- date
  • Include source citations
  • More reliable than generic answers

How RAG Works​

Customer asks: "What's your refund policy?"
 ↓
AI searches your documents
 ↓
Finds relevant information in "Returns Policy.pdf"
 ↓
AI generates answer using that information
 ↓
Customer sees: "Our refund policy allows returns within 30 days..."
 + Source: Returns Policy.pdf, page 3

AI Providers & Configuration​

Axie Studio supports multiple AI providers. Choose the one that works best for you:

OpenAI​

  • Best for: High- quality, general- purpose responses with large context windows
  • Models:
  • GPT- 5 (400k context window) - Latest model, best for complex documents
  • GPT- 4o (128k context window) - Optimized for quality and speed
  • GPT- 4o Mini (128k context window) - Cost- effective option
  • GPT- 4 Turbo (128k context window) - High- performance option
  • When to use: When you need the best quality answers and large context windows

Cloudflare Workers AI​

  • Best for: Fast, cost- effective responses with serverless architecture
  • Models:
  • Llama 3.1 8B Instruct (8k context window) - Fast and efficient
  • Llama 3.1 70B Instruct (8k context window) - Higher quality, larger model
  • When to use: When speed and cost matter, or for serverless deployments

DeepSeek​

  • Best for: Technical and coding questions
  • Models:
  • DeepSeek Chat (32k context window) - General purpose
  • DeepSeek Coder (16k context window) - Optimized for code
  • When to use: For technical documentation and coding- related queries

Choosing an AI Provider​

  1. Go to Dashboard -> Knowledge Base
  2. Select your knowledge base
  3. Go to AI Settings
  4. Choose your provider
  5. Select a model

LLM Provider Configuration

Configure your AI provider and model settings to control how your AI agent generates responses. Multiple providers are supported including OpenAI (GPT- 5, GPT- 4o), Cloudflare Workers AI (Llama 3.1), and DeepSeek (Chat/Coder).

Temperature Settings​

Control how creative AI responses are:

  • 0.0 - 0.3: Very factual, consistent answers
  • 0.4 - 0.7: Balanced (recommended)
  • 0.8 - 1.0: More creative, varied answers
  • 1.1 - 2.0: Very creative (use carefully)

Recommended: Start with 0.7 for balanced responses.

Response Length​

Control how long answers are:

  • Short (100- 200 tokens): Brief, concise answers
  • Medium (200- 500 tokens): Detailed answers (recommended)
  • Long (500+ tokens): Very detailed, comprehensive answers

Recommended: Start with 500 tokens for detailed but not overwhelming answers.

Document Processing & Search​

How Documents Are Processed​

When you upload a document:

  1. Text Extraction: Text is extracted from PDF/DOCX/TXT
  2. Chunking: Document is split into searchable pieces
  3. Embedding: Each chunk gets a searchable "fingerprint"
  4. Storage: Chunks are stored in a searchable database

Embeddings Configuration​

Axie Studio uses Voyage AI for document embeddings:

  • Model: Voyage AI 3 Large (voyage- 3- large)
  • Purpose: Converts document chunks into searchable vector embeddings
  • Quality: High- quality embeddings for accurate semantic search

Chunking Explained​

Documents are split into "chunks" for better search:

  • Chunk Size: Default 512 tokens (about 400 words)
  • Overlap: 50 tokens between chunks (ensures context)
  • Why: Smaller chunks = more precise search results
info

You can adjust chunk size in advanced settings if needed.

Document Parsers​

Choose how documents are processed:

  • Default Parser (Serverless): Handles PDF, TXT, DOCX files automatically
  • Reducto AI: Advanced parsing for complex documents (requires API key)

Vector Store​

Axie Studio uses ChromaDB Cloud (serverless) for vector storage:

  • Type: ChromaDB Cloud - Fully managed, serverless
  • Benefits: No infrastructure to manage, automatic scaling
  • Configuration: Requires ChromaDB Cloud API key, Tenant ID, and Database name

How Search Works​

When a customer asks a question:

  1. Query Embedding: Question is converted to a searchable format
  2. Similarity Search: System finds most relevant document chunks
  3. Top- K Retrieval: Gets top 5 most relevant chunks (default)
  4. Context Building: Combines chunks into context
  5. Answer Generation: AI generates answer using context

Top- K Settings​

Control how many document chunks to use:

  • Lower (3- 5): More focused, faster
  • Higher (7- 10): More comprehensive, slower

Recommended: Start with 5 chunks for balanced results.

Advanced Settings​

System Prompt​

Customize how AI responds:

  • Set the tone (formal, casual, friendly)
  • Define response style
  • Add specific instructions

Chunk Settings​

Control how documents are split:

  • Chunk size (default: 512 tokens)
  • Overlap between chunks (default: 50 tokens)
  • Optimize for your content type

Retrieval Settings​

Control search behavior:

  • Number of chunks to retrieve (Top- K)
  • Similarity threshold
  • Maximum context length

Source Citations​

Every answer includes where information came from:

Answer: "Our refund policy allows returns within 30 days..."

Source: Returns Policy.pdf, page 3

This helps:

  • Build customer trust
  • Verify information accuracy
  • Find original documents
  • Improve documentation

Troubleshooting​

AI Not Finding Information?​

Try:

  • Make sure information is in your documents
  • Check if documents processed successfully
  • Try rephrasing the question
  • Increase Top- K value

Answers Not Accurate?​

Improve:

  • Make documents more specific
  • Add more examples
  • Adjust system prompt
  • Try different AI models
  • Lower temperature for more factual answers

Slow Responses?​

Optimize:

  • Use faster AI models (Cloudflare)
  • Reduce Top- K value
  • Check document size
  • Optimize chunk settings

Processing Failed?​

Fix:

  • Check file format (PDF, DOCX, TXT only)
  • Verify file size (max 10MB)
  • Try a different file
  • Contact support if issues persist

Next Steps​

Ready to add documents? Go to Dashboard Knowledge Base and create your first knowledge base!