Skip to content

Connecting to Assistants

Learn how to link your knowledge bases to assistants for intelligent, context-aware responses.

Prerequisites

  • Store created
  • Files uploaded and processed
  • Assistant configured

Connecting a knowledge base enables contextual answers based on your documents and reduces hallucination by grounding responses in real data.

Connecting a Store

Method 1: From Assistant Detail View

  1. Go to Assistants
  2. Click Manage Assistant on your desired assistant
  3. Scroll to the Stores section
  4. Click Add Store
  5. Select the store from the dropdown
  6. The store is now connected

Method 2: From Chat Interface

  1. Go to Chat
  2. Select your assistant
  3. In the middle panel, click Add Store
  4. Select the store to connect

This is useful for quick testing without leaving the chat.

Managing Connections

View Connected Stores

In the assistant detail view, the Stores section shows:

  • List of connected stores
  • Number of connections (badge)
  • Quick disconnect option

Disconnect a Store

  1. Go to the assistant's Stores section
  2. Click the disconnect icon next to the store
  3. The store is removed from the assistant

Non-Destructive

Disconnecting doesn't delete the store or its files. You can reconnect anytime.

How It Works

When a user asks a question, the system searches your connected stores for relevant information and uses it to generate accurate responses.

Advanced: Query Flow & Retrieval Settings

Query Flow: 1. User asks a question (e.g., "What's your return policy?") 2. Question is converted to a vector embedding 3. System searches all connected stores for similar chunks 4. Top N chunks (default: 5) are retrieved based on similarity 5. Retrieved chunks provide context for the AI's response 6. User receives a response grounded in your documentation

Retrieval Settings (configured in assistant's advanced options):

Setting Default When to Adjust
Max Results 5 Increase to 8-10 if answers lack detail, decrease to 3-4 if responses feel scattered
Score Threshold 0.7 Raise to 0.8 if getting irrelevant results, lower to 0.6 if missing matches

Multiple Stores

Connecting Multiple Stores

You can connect multiple stores to a single assistant. The assistant will search across all connected stores.

Use cases: - Different content categories - Separate departments contributing knowledge - Staged content updates

Best Practices for Multiple Stores

  1. Keep stores focused: One topic per store
  2. Avoid duplicates: Don't upload the same content to multiple stores
  3. Consider relevance: Only connect stores relevant to the assistant's purpose

Testing Your Connection

  1. Go to Chat
  2. Select your assistant with the connected store
  3. Ask a question specific to your uploaded content
  4. The response should reference information from your documents

Good test questions: - Questions only answerable from your documents - Specific facts, figures, or policies

Debugging Poor Results

If the assistant doesn't use your knowledge base:

  1. Check connection: Verify the store is listed in the assistant
  2. Verify processing: Ensure you clicked "Finish Upload" after uploading
  3. Test search: Ask very specific questions matching document content

Best Practices

Quality Content: - Well-structured documents with clear headings - Complete sentences with context - Focused, relevant information - Regular updates to keep content current

Clear Assistant Instructions:

Your assistant's instructions should explicitly mention using the knowledge base.

Good Instruction Example

You are a customer support assistant for [Company Name].

When answering questions, always search your knowledge base first
and base your answers on the information found there.

If you cannot find information about a topic in your knowledge base,
say: "I don't have specific information about that in my documentation.
Let me connect you with a human agent who can help."

Never make up policies or product details. Only use information from
your knowledge base.

Test Thoroughly: - Questions you know are answered in your docs - Questions with tricky phrasing - Questions that aren't in your docs (to test fallback behavior)

Advanced: Store Organization Strategies

Choose an organization strategy that fits your use case:

By topic (most common):

Assistant: "Product Support Bot"
Stores:
- Product Specifications (40 MB)
- Troubleshooting Guides (25 MB)
- Warranty Information (5 MB)

Why: Different document types, easy to update specific areas

By update frequency:

Assistant: "Company Information Bot"
Stores:
- Static Content (20 MB) - Rarely changes
- Dynamic Content (30 MB) - Changes weekly
- Seasonal Content (10 MB) - Updated quarterly

Why: Easy to update frequently-changing content without touching stable information

By product line:

Assistant: "Multi-Product Support"
Stores:
- Widget Pro Line (35 MB)
- Widget Lite Line (25 MB)
- Accessories (15 MB)

Why: Clear separation, can connect/disconnect stores based on context

Advanced: Prompt Engineering for RAG

Enhance your assistant's use of RAG with these prompt techniques:

Confidence calibration:

If the information in your knowledge base is unclear or
contradictory, acknowledge this and offer to escalate to
a human agent.

Handling gaps:

If a user's question is partially covered in your knowledge base,
answer what you can from your documents and clearly indicate
what information you don't have available.

Multi-store awareness:

You have access to multiple knowledge bases covering different
topics. Search all of them to find relevant information.

Troubleshooting

Assistant ignores knowledge base: - Verify store is connected - Check processing completed - Add explicit instructions to use the knowledge base - Test with questions that match document content exactly

Incorrect information retrieved: - Review document content for accuracy - Remove outdated or incorrect documents - Check for duplicate/conflicting information

Slow responses: - Reduce number of connected stores - Optimize document sizes - Adjust retrieval settings (fewer results)

No results found: - Verify documents were processed (Finish Upload) - Check document format is supported - Ensure content is text-based (not images)

Next Steps

Test Your Assistant