Skip to main content

Knowledge Bases: Adding Specific Information to Your AI Agents

Knowledge Bases allow you to add your company's specific information—like products, policies, and procedures to your AI agents. This page covers both Fast Access and Vector Knowledge Bases, setup procedures, optimization strategies, and troubleshooting.

Prerequisites​

  • Active FlameTree account with agent creation permissions
  • Documents in supported formats (PDF, DOCX, TXT, HTML, MD, JSON, XLSX)
  • For Notion integration: Notion workspace with appropriate permissions
  • For web sources: Publicly accessible websites

What Are Knowledge Bases?​

Your AI agent initially only knows what was included during its training, which doesn't include specific information about your company, products, or services. Knowledge Bases solve this by adding your specific information to the agent's available knowledge.

FlameTree provides two methods to add knowledge to your agents:

  • Fast Access Knowledge Base - Essential information that's immediately available in every conversation
  • Vector Knowledge Base - Large-scale searchable documentation using advanced retrieval technology

Choose Your Knowledge Base Type​

NeedRecommended TypeSetup Time
FAQ, company policies, contact infoFast Access Knowledge Base2 minutes
Product catalogs, technical documentationVector Knowledge Base10 minutes
Both essential info AND detailed docsUse both types together15 minutes

💡 Note: You can use both types together. Fast Access provides instant information while Vector Knowledge Base handles complex queries.

Knowledge Base Types Comparison​

When to Use Each Type​

Fast Access Knowledge BaseVector Knowledge Base
✅ FAQ and essential policies✅ Product catalogs and specifications
✅ Contact information✅ Technical documentation
✅ Company operating hours✅ Large document libraries
✅ Basic product information✅ Historical data and records

Key Differences​

Setup & Management:

  • Fast Access: Simple text field in agent settings → Save → Restart agent
  • Vector Knowledge Base: Create → Upload sources → Build → Configure tools → Restart agent

Performance & Cost:

  • Fast Access: Instant responses, but makes all conversations slower and more expensive
  • Vector Knowledge Base: Slight search delay, but only costs when information is retrieved

Size Limitations:

  • Fast Access: ~10,000 characters maximum (must fit in model prompt)
  • Vector Knowledge Base: Virtually unlimited (recommended: files under 300 MB each)

Technical Requirements​

FeatureFast AccessVector Knowledge Base
Tool ConfigurationNone requiredRequires SingleSearchFAQ tool
File FormatsText only (manual entry)PDF, DOCX, TXT, HTML, MD, JSON, XLSX
UpdatesEdit text and restart agentUpload new files → Rebuild → Restart agent
Search MethodAlways availableRAG-based vector search

Content Capabilities​

Fast Access Knowledge Base:

  • ✅ Text-only, manually entered content
  • ✅ Always embedded in every conversation
  • ❌ Limited formatting options
  • ❌ Cannot handle large amounts of information

Vector Knowledge Base:

  • ✅ Multiple file formats supported
  • ✅ Advanced search capabilities
  • ✅ Scalable for large datasets
  • ❌ Text-only sources (images/diagrams not indexed)
  • ❌ Requires rebuilding when updated

Quick Setup Guide​

Fast Access Knowledge Base (2 minutes)​

  1. Go to your agent settings
  2. Add text to the "Fast Access Knowledge Base" field
  3. Save and restart your agent

Vector Knowledge Base (10 minutes)​

  1. Navigate to "Knowledge Base" in the left menu
  2. Click "Create New Knowledge Base"
  3. Upload your documents or add web sources
  4. Click "Build" and wait for processing
  5. Connect to your agent and restart

Setting Up Vector Knowledge Base (RAG)​

🧐 How Vector Knowledge Base Works?

The Vector Knowledge Base uses Retrieval Augmented Generation (RAG) through a five-step process:

  1. Document Processing

    • Documents are broken into small chunks based on content structure (paragraphs, sections)
    • Configurable chunk size determines the size of each piece
    • Overlap settings preserve context between adjacent chunks

  2. Vectorization

    • Each chunk is converted to a mathematical representation (vector) using embedding models
    • These vectors capture the semantic meaning of the text content
    • The embedding model transforms text into numerical representations for similarity matching

  3. Storage

    • Chunks and their vectors are stored in a searchable database optimized for similarity matching
    • Two storage options are available:
      • Faiss (default) – In-memory vector store that is simple and efficient for most use cases
      • Qdrant (experimental) – A high-performance vector database offering faster indexing and search, suitable for larger or more demanding knowledge bases

  4. Retrieval

    • When the agent needs information, the user query is converted to a vector
    • The system finds the most similar chunks using cosine similarity or similar metrics
    • The most relevant chunks are returned to the agent based on similarity scores

  5. Response Generation

    • The agent uses the found information along with its training to generate contextually appropriate responses
    • Retrieved chunks provide specific knowledge while the agent's training handles language generation and reasoning

Step 1: Create Your Knowledge Base​

  1. Navigate to Knowledge Base

    • Open the left menu and click "Knowledge Base"
    • Click "Create New Knowledge Base"
    • Provide a descriptive name and description

  2. Choose Your Configuration

    • Start with default settings for your first knowledge base
    • You can optimize these later based on performance

Step 2: Add Your Information Sources​

File Sources - Upload Documents

Supported Formats: PDF, DOCX, TXT, HTML, Markdown (.md), JSON, Excel (.xlsx)

Setup Process:

  1. Click "Add Source" → "File"
  2. Upload your file
  3. Select the correct format if not auto-detected
  4. System automatically parses content

Best Practices:

  • Ensure files are well-formatted with clear structure
  • Remove irrelevant headers/footers from PDFs
  • Use consistent formatting across documents
  • Keep files under 100-300 MB for best performance
Web Sources - Crawl Websites

Two Options Available:

  • Static Web Pages: Extract content from a single webpage
  • Dynamic Web Crawling: Automatically crawl entire website following internal links

Setup Process:

  1. Click "Add Source" → "Web"
  2. Choose Static or Dynamic crawling
  3. Provide the starting URL
  4. For dynamic crawling, configure crawling depth
  5. Select "Parse all site pages" to crawl the entire site recursively

⚠️ Warning: If a site is protected against scraping, manually save pages as HTML and upload as files instead.

Notion Integration

Detailed Setup Process:

1. Create a Notion Integration

  • Go to Notion: Settings → Connections → Developer / Manage Integrations
  • Click Create new integration
  • Enter a name (e.g., "Flametree Training")
  • Select your workspace and click Save
  • Copy the generated Integration Secret (API key)

2. Connect Integration to Your Page

  • Open the Notion page you want to include
  • Click three dots (⋯) → Connections
  • Select your newly created integration
  • Click Confirm

3. Configure in FlameTree

  • Enter your Notion Token (from step 1)
  • Enter Notion Page URL (everything before the # symbol)
Manual Sources - Add Custom Content

Purpose: Add specific information that isn't available in document form

Setup Process:

  1. Click "Add Source" → "Manual"
  2. Enter a descriptive title
  3. Add your content
  4. Specify any relevant metadata
  5. Save the source

Step 3: Configure Advanced Parameters​

Chunk Size (Default: 1000 characters)

What It Does: Determines how large each piece of information will be when your documents are processed.

Smaller Chunks (500-800 characters):

  • ✅ More precise answers to specific questions
  • ✅ Better for technical documentation and FAQs
  • ❌ May miss broader context

Larger Chunks (1200-1500 characters):

  • ✅ Preserves more context and relationships
  • ✅ Better for narrative content and marketing materials
  • ❌ May include irrelevant information

Recommendation: Start with 1000 characters and adjust based on your content type and testing results.

Chunk Overlap (Default: 20 characters)

Purpose: Prevents important information from being split across chunks

Settings:

  • Low overlap (10-20 characters): Standard for most content
  • High overlap (50-100 characters): Better for technical content with complex relationships

Guidelines: Helps preserve meaning across boundaries. ~20 characters is usually sufficient, but this depends on your text structure.

Top-K Results (Default: 5)

Purpose: Number of relevant chunks returned for each search

Lower Values (3-4):

  • Pros: More focused responses, less noise
  • Cons: May miss relevant information
  • Best for: Simple queries, specific information needs

Higher Values (7-10):

  • Pros: More comprehensive information, better coverage
  • Cons: May include irrelevant data, longer processing time
  • Best for: Complex queries, research-type questions

Guidelines: A higher number increases context but can introduce noise.

Embedder Model

Purpose: The model used to generate vector representations

Options: Several models are supported (e.g., OpenAI-compatible)

Selection: Choose the one that best fits your content volume and budget

Importance: Accurate embeddings are key for effective search

Storage Type

Purpose: Specifies the vector storage backend for the knowledge base

Options:

  • Faiss (default) – In-memory vector store that is simple and efficient for most use cases
  • Qdrant (experimental) – A high-performance vector database offering faster indexing and search, suitable for larger or more demanding knowledge bases

Step 4: Build and Test​

  1. Review Your Sources

    • Confirm all documents are uploaded correctly
    • Verify configuration settings

  2. Build the Knowledge Base

    • Click "Build" to start processing
    • Wait for processing to complete (status shows progress)
    • Knowledge base status changes to "Ready"

  3. Test Your Knowledge Base

    • Use the built-in search tools to test queries
    • Ask questions your users would actually ask

⚠️ Warning: The building process can take several minutes for large knowledge bases.

Connecting Knowledge Base to Your Agent​

  1. Link Knowledge Base to Agent

    • Go to your agent configuration
    • In the "Knowledge Base" section, select your created knowledge base
    • Save the agent configuration

  2. Configure Workflow Tools For Vector Knowledge Base to work, your agent's workflow must include the search tool:

    available_tools:
      SingleStatefulOutboundAgent:
        - SendChatMessage
        - SingleSearchFAQ  # Required for knowledge base search
        - FinishSession

  3. Restart Your Agent

    • Stop the agent if it's currently running
    • Start the agent again to load the new configuration
    • Test with sample questions

💡 Note: Agents only load knowledge bases when they start. Always restart your agent after making knowledge base changes.

Real-World Example: Bank Customer Service​

The Challenge​

A bank wants their AI agent to answer questions about credit cards, loans, and deposit accounts with accurate, up-to-date information.

The Solution​

Documents Added:

  • PDF: Credit Card Terms and Conditions (450 pages)
  • PDF: Personal Loan Application Guide (75 pages)
  • Excel: Current Interest Rates and Fees (updated monthly)
  • Web: Public rate information pages
  • Manual: Recent policy updates and announcements

Configuration Used:

Chunk Size: 800 characters (for precise financial information)
Chunk Overlap: 30 characters (preserve rate relationships)
Top-K Results: 6 (comprehensive coverage for financial queries)
Embedder: Multilingual (support multiple languages)

Sample Interaction​

User Question: "What's the cash withdrawal fee for the Visa Gold card?"

System Process:

  1. Converts query to vector representation
  2. Searches for similar chunks about "cash withdrawal fee" and "Visa Gold"
  3. Finds relevant sections from credit card terms document
  4. Returns specific fee structure, limits, and conditions

Agent Response: "The cash withdrawal fee for the Visa Gold card is 2% with a minimum of $3 per transaction. This applies to both domestic and international ATM withdrawals. For ATMs within our network, the fee is reduced to $1 per transaction."

Results​

  • 95% accuracy on fee and rate questions
  • Reduced customer service calls by 40%
  • Consistent, up-to-date information across all channels

Managing and Maintaining Knowledge Bases​

Adding New Information​

Process:

  1. Navigate to your knowledge base
  2. Click "Add Source" to include new documents
  3. Knowledge base status changes to "Modified"
  4. Click "Build" to rebuild with new information
  5. Restart your agent to load updates

Updating Existing Information​

For Document Sources:

  1. Remove outdated sources
  2. Upload updated versions
  3. Rebuild knowledge base
  4. Test to ensure accuracy

For Web Sources:

  1. Web pages are re-crawled during rebuild
  2. Simply click "Build" to get latest content
  3. Restart agent to apply updates

For Notion Sources:

  1. Update content in Notion
  2. Rebuild knowledge base in FlameTree
  3. Restart agent

⚠️ Critical: Always restart your agent after rebuilding a knowledge base. Changes won't take effect until the agent is restarted.

Regular Maintenance Schedule​

Maintenance Tasks by Frequency​

FrequencyTasksPurpose
Monthly• Review and update frequently changing information (prices, policies)
• Test agent responses for accuracy
• Add new sources as needed		Keep information current and relevant
Quarterly• Comprehensive review of all content for accuracy
• Performance optimization based on usage patterns
• Remove outdated or redundant information		Maintain quality and performance
Annual• Major restructuring and reorganization
• Review and update configuration parameters
• Archive and replace old documentation		Strategic improvements and cleanup

Monitoring Framework​

Monitoring CategoryWhat to TrackAction Items
Content Quality• Frequently requested but not found information
• Agent response accuracy
• User feedback about missing information		• Add missing content
• Update incorrect information
• Create new documentation sections
Performance• Response times and accuracy
• Search result relevance
• Usage patterns and trends		• Adjust chunk size and overlap
• Optimize Top-K settings
• Fine-tune embedding models
Version Control• Knowledge base versions and changes
• Documentation of updates
• Backup maintenance		• Track all modifications
• Document reasons for changes
• Maintain rollback capabilities

Maintenance Checklist​

Monthly Review Checklist:

  • Update pricing information and policies
  • Test 10-15 common user questions for accuracy
  • Review agent logs for "information not found" responses
  • Add any new documentation or sources
  • Check for broken web sources or outdated links

Quarterly Optimization Checklist:

  • Analyze search performance metrics
  • Review and adjust chunk size/overlap settings
  • Remove duplicate or redundant information
  • Update embedding models if newer versions available
  • Conduct comprehensive accuracy testing

Annual Strategic Review Checklist:

  • Evaluate overall knowledge base structure
  • Consider splitting large knowledge bases by topic
  • Review and update configuration parameters
  • Archive outdated documentation
  • Plan knowledge base expansion or consolidation

Testing and Optimization​

Built-in Testing Tools​

Your knowledge base includes powerful testing tools to help you optimize performance:

ToolPurposeHow to UseExample
Text SearchFind chunks containing specific wordsType exact phrases you want to findSearch "refund policy" to verify indexing
Vector SearchTest semantic similarity searchEnter questions users would askSearch "return item" to see relevant chunks

Performance Optimization Process​

1. Start with Default Settings

  • Build your knowledge base with standard configuration
  • Test with representative questions from your users

2. Analyze Search Results

  • Use vector search to see which chunks are retrieved
  • Check similarity scores (higher scores = more relevant)
  • Identify gaps in coverage or irrelevant results

3. Adjust Parameters Based on Results

IssueSolution
Incomplete answersIncrease Top-K or chunk size
Too much irrelevant infoDecrease Top-K or chunk size
Missing contextIncrease chunk overlap
Answers too genericDecrease chunk size, improve source quality

4. Test in Agent Playground

  • Go to Playground tab
  • Select your agent
  • Ask realistic questions based on your knowledge base content
  • Verify the agent references actual data from your documents

5. Iterate and Improve

  • Make incremental changes
  • Rebuild knowledge base after each change
  • Restart agent and test again
  • Continue until satisfied with performance

Optimization Guidelines by Content Type​

Technical Documentation:

Chunk Size: 500-800 characters
Chunk Overlap: 30-50 characters
Top-K Results: 7-10
Reasoning: Technical info needs precision and comprehensive coverage

Marketing Content:

Chunk Size: 1200-1500 characters
Chunk Overlap: 20-30 characters
Top-K Results: 5-6
Reasoning: Marketing benefits from context and narrative flow

Mixed Content:

Chunk Size: 1000 characters (default)
Chunk Overlap: 20 characters (default)
Top-K Results: 5 (default)
Reasoning: Start with defaults, adjust based on testing

Best Practices​

Content Organization​

Structure Your Sources Logically:

✅ Good Organization:
Products/
  ├── Credit_Cards/
  │   ├── Terms_and_Conditions.pdf
  │   ├── Fee_Schedule.pdf
  │   └── Application_Process.pdf
  ├── Loans/
  │   ├── Personal_Loans.pdf
  │   └── Mortgage_Information.pdf
  └── Support/
      ├── FAQ.pdf
      └── Contact_Information.pdf

Content Quality Guidelines​

  • Remove duplicates: Eliminate redundant content across documents including outdated versions of documents
  • Use consistent terminology: Standardize language across all sources and maintain consistent terms across all documents
  • Include customer language: Use terms customers actually use, not just internal jargon. Include common misspellings or variations, add synonyms for technical terms.
✅ Include both:
- Technical term: "Account reconciliation"
- Customer term: "Balance check" or "Account verification"
  • Structure clearly: Use headings, bullet points, and consistent formatting
  • Prepare clean, well-formatted content: Remove redundant or outdated information, use clear headings and structure, include relevant examples and use cases, maintain consistent formatting standards

Parameter Optimization​

Systematic Optimization Process​

  1. Start with Defaults: Begin with standard settings
  2. Test with Real Queries: Use questions your customers would ask
  3. Analyze Results: Use built-in search tools to examine retrieved information
  4. Adjust Incrementally: Make small parameter changes
  5. Validate Improvements: Test to confirm better performance
  6. Document Changes: Keep track of optimization decisions

Common Optimization Patterns​

For Technical Documentation:

Chunk Size: 600-800 characters
Chunk Overlap: 40 characters
Top-K Results: 7-8
Reasoning: Technical content needs precision and comprehensive coverage

For Marketing Content:

Chunk Size: 1200-1400 characters  
Chunk Overlap: 25 characters
Top-K Results: 5-6
Reasoning: Marketing content benefits from context and narrative flow

Size Management Strategies​

For Large Organizations:

  • Create separate knowledge bases by topic (Products, Support, Policies)
  • Use multiple specialized agents for different domains
  • Keep individual knowledge bases focused and manageable

File Size Guidelines:

  • Keep files under 100-300 MB for optimal performance
  • Split very large documents into logical sections

Common Issues & Solutions​

Agent Can't Find Information That Should Be Available

Symptoms: Agent says information is not available when it should be

Solutions:

  1. Check Source Processing: Ensure documents were processed correctly
  2. Test Search Tools: Use vector search to see if information is findable
  3. Adjust Query Phrasing: Try different ways of asking the same question
  4. Increase Top-K: More chunks returned may include the needed information
  5. Review Chunk Size: Information might be split across chunks
Information is Outdated or Incorrect

Solutions:

  1. Update Sources: Replace outdated documents with current versions
  2. Rebuild Knowledge Base: Process updated information
  3. Restart Agents: Ensure agents load the updated knowledge base
  4. Remove Old Sources: Delete outdated documents from the knowledge base
Responses Too Generic or Unhelpful

Solutions:

  1. Add More Specific Information: Include detailed documentation
  2. Improve Content Structure: Use clear headings and organization
  3. Increase Chunk Overlap: Preserve more context between chunks
  4. Add Examples: Include specific examples and use cases
Too Much Irrelevant Information

Symptoms: Agent includes unrelated information in responses

Solutions:

  1. Decrease Top-K: Return fewer chunks per search
  2. Improve Source Quality: Remove or clean up noisy documents
  3. Smaller Chunk Size: More precise information retrieval
  4. Better Source Organization: Separate different types of information
Knowledge Base Access Issues

Symptoms: Agent cannot access or query the knowledge base

Diagnostic Checklist:

  • Knowledge base status is "Ready" (not "Modified")
  • Knowledge base is connected to agent in agent configuration
  • SingleSearchFAQ tool is listed in workflow configuration
  • Agent has been restarted after configuration changes
File Upload and Processing Issues

Symptoms: Uploaded files do not appear in the Knowledge Base interface, are not indexed, or the base remains in Modified status

Solution: Check the file format and convert it if necessary.

Supported Formats:

FormatPurposeNotes
.pdfDocuments, presentations, reportsFull-text parsing supported
.docxMicrosoft Word documents
.txtPlain textUTF-8 encoding recommended
.mdMarkdown documentationIdeal for FAQs and instructions
.htmlExported web pages and articlesNo scripts, clean HTML only
.jsonlStructured records (JSON Lines)Commonly used for FAQs
.xlsxExcel spreadsheets

If your file format is not supported, convert it to one of the formats listed above.

Web Source Not Working

Common reasons why a Web source may fail to work:

  1. The site is protected against scraping - Flametree parser cannot retrieve the page content
  2. The site is inaccessible from Flametree servers - May be IP-blocked or geographically restricted

Solutions:

  • Manually save the page as .html and upload it to the KB as a file
  • In Advanced settings of WEB source, decrease number of tasks per minute to avoid blocking for too frequent requests

FAQ​

How do I connect a knowledge base to my AI agent?

Four mandatory steps are required:

Step 1. Create a Knowledge Base

  • Go to the Knowledge Bases tab
  • Click the Create Knowledge Base button
  • Specify: name, source type (file, Notion, Web, etc.), settings: chunk_size, chunk_overlap, embedding_model, top_k
  • After uploading all sources, click Build

Step 2. Link it to the AI agent

  • Go to the desired AI agent's page
  • In the Database section, select the previously created knowledge base

Step 3. Add tools to the Workflow

  • The AI agent cannot access the knowledge base unless the appropriate tool (SingleSearchFAQ) is explicitly declared in the workflow

Example configuration:

available_tools:
  SingleStatefulOutboundAgent:
    - SendChatMessage
    - FinishSession
    - SingleSearchFAQ

Step 4. Restart Your Agent

  • Stop the agent if it's currently running
  • Start the agent again to load the new configuration

❗️ Important: Fast Access KB does not require a tool call (embedded directly into the prompt), but RAG (vector knowledge base) requires an explicit tool call via the workflow.

What is an embedding model and why is multilingual important?

An Embedding Model converts text from a knowledge base into vector representations used to find the most relevant text fragments (chunks) when searching for answers.

How it works:

  • Each uploaded document is split into chunks
  • Each chunk is processed by the embedding model and transformed into a vector
  • During interaction, user queries are converted to vectors and matched against stored chunks
  • This is the core of the RAG (retrieval-augmented generation) mechanism

Multilingual embedding models:

  • Support multiple languages simultaneously
  • Allow documents in different languages (e.g., English and Russian)
  • Enable users to ask questions in different languages

💡 Note: Essential if your knowledge base contains materials in multiple languages or you want the agent to respond to clients in different languages.

What does the "Build" button do and when should I use it?

The Build button processes and indexes your knowledge base, making it searchable and usable for AI agents.

What happens during Build:

  1. Source processing - Reads all uploaded documents and data
  2. Chunking - Splits documents into fragments based on chunk_size and chunk_overlap
  3. Vectorization - Processes chunks through the embedding model
  4. Indexing - Stores vectors in the search database
  5. Activation - Makes the knowledge base ready for agent use

When to click Build:

  • After creating a new knowledge base
  • After uploading new documents
  • After changing configuration parameters (embedding_model, chunk_size, chunk_overlap, top_k)
  • Anytime status shows "Modified"

⚠️ Warning: Building can take several minutes for large knowledge bases. Don't close your browser during the process.

What does "Modified" status mean and how do I fix it?

The "Modified" status means changes have been made but not yet applied to the knowledge base.

What causes Modified status:

  • Sources added or removed (files, links, Notion, etc.)
  • Configuration parameters changed (embedding_model, chunk_size, chunk_overlap, top_k)
  • Language settings or Multilingual option modified
  • Settings saved but Build not executed

How to fix: Click the "Build" button on the knowledge base card to trigger re-processing.

What happens if you don't fix it:

  • Agent cannot use the changes
  • Search responses rely on the previous version
  • Some sources may be ignored entirely

❗️ Important: Always click "Build" after making changes, then restart your agent.

What are the size limitations for knowledge bases?

Vector Knowledge Base:

  • Recommended: Files under 100-300 MB each
  • Maximum: 500 MB per file (larger files may fail or crash)
  • Total size: No strict limit, but 10GB+ datasets may affect performance

💡 Note: Avoid massive all-in-one .pdf compilations.

Fast Access Knowledge Base:

  • Recommended: 5,000-10,000 characters maximum
  • Constraint: Must fit entirely in the model prompt
  • Content: Text only, manually entered

💡 Note: For large datasets, split into separate logical knowledge bases rather than using massive all-in-one files.

What are Top-K, Chunk Size, and Chunk Overlap parameters?

These parameters control how documents are processed and which information is retrieved during searches.

Chunk Size - Size of each text fragment after document splitting

  • Smaller (500-800): More precise answers, better for technical docs
  • Larger (1200-1500): More context, better for narrative content
  • Default: 1000 characters

Chunk Overlap - Number of overlapping characters between adjacent chunks

  • Purpose: Preserves context across chunk boundaries
  • Typical range: 10-100 characters
  • Default: 20 characters

Top-K - Number of relevant chunks returned for each search

  • Lower (3-4): More focused responses, less noise
  • Higher (7-10): More comprehensive coverage, potential irrelevant data
  • Default: 5 chunks

💡 Note: Start with defaults and adjust based on testing results with your specific content.

What's the difference between Fast Access and Vector Knowledge Bases?
FeatureFast Access KBVector Knowledge Base
PurposeQuick reference, FAQsLarge-scale documentation
How it worksEmbedded directly in promptVector search (RAG)
Size limit~10,000 charactersVirtually unlimited
Tool requiredNoneSingleSearchFAQ
Setup time2 minutes10 minutes
File supportText only (manual entry)PDF, DOCX, TXT, HTML, MD, JSON, XLSX
Cost impactMakes all conversations more expensiveOnly costs when information is retrieved

💡 Note: You can use both types together for optimal performance.

Which characters are not allowed in Fast Access Knowledge Base?

The hash symbol (#) has special meaning in Fast Access Knowledge Base for structuring content and marking headings.

Restriction:

  • Do not use fewer than 4 consecutive hash symbols (####) at the start of a line
  • Lines beginning with #, ##, or ### may be misinterpreted as malformed structure

Recommendation: Use #### or more hash symbols for headings to avoid parsing errors.

What language should I use in my knowledge bases?

Best Practice:

Use the same language that your AI agent is expected to respond in.

This applies to all source types:

  • Documents (PDF, Word, HTML, etc.)
  • Web pages
  • Fast Access blocks

For Multilingual Use Cases:

  1. Enable "Multilingual" when creating the knowledge base
  2. Configure the AI agent to detect input language and respond accordingly
  3. Upload documents in multiple languages as needed

💡 Note: Multilingual embedding models support multiple languages simultaneously and allow cross-language querying.

Why is my web source not working?

Common Causes:

  1. Site protected against scraping - Flametree parser cannot retrieve content
  2. Site inaccessible - May be IP-blocked or geographically restricted from Flametree servers

Solutions:

  • Manually save the page as .html and upload it as a file source
  • In Advanced settings of web source, decrease tasks per minute to avoid rate limiting
  • Check if the site requires authentication or has bot protection

Checklist​

  • Restart the bot after any changes - Applies to: workflow, skills, environment variables, knowledge bases, etc.

  • Update the knowledge base using the Build button, then restart the bot - This recalculates embedding vectors and applies new sources.

  • Monitor the knowledge base status - If the status is Modified — click Build, or changes will not take effect.

  • Use only supported file formats - Convert unsupported formats before uploading.

  • Limit the size of uploaded files - Recommended: up to 100–300 MB. Files over 500 MB may fail to upload or crash.

  • Don't forget to declare all required available_tools - The AI agent won't be able to use the knowledge base if the search tool is missing in the current state.

  • Avoid overloading the prompt with Fast Access KB - Recommended size: up to 5,000–10,000 tokens (roughly 5–10 pages of text).