Knowledge Engine

The Knowledge Engine enables RAG (Retrieval-Augmented Generation) for your AI applications. Upload documents, and the LLM automatically retrieves relevant context to answer questions with your private data.

Feature Access Required: Knowledge Engine requires whitelisted access.

To request access, email sales@demeterics.com with:

• Subject: "Feature Access Request"
• Feature name: "Knowledge Engine"

Overview

The Knowledge Engine transforms your documents into a searchable knowledge base that LLMs can query. Instead of relying solely on training data, the LLM retrieves relevant passages from your documents to provide accurate, grounded responses.

Key Benefits:

• Grounded Responses: Answers based on your actual documents, not hallucinated content
• Private Data: Your documents stay in your Demeterics account, not sent to LLM training
• Automatic Indexing: Upload documents and they're immediately searchable
• Multi-Format Support: PDF, Word, Markdown, HTML, and plain text

Architecture

User Query → LLM (with Knowledge Tools)
                    ↓
            search_knowledge (vector search)
            find_documents (document discovery)
            get_summary (topic overview)
            get_content (full document)
                    ↓
            Relevant context injected
                    ↓
            LLM generates grounded response

The Knowledge Engine provides four tools that the LLM can call during conversations:

Getting Started

Step 1: Create a Knowledge Project

1. Go to Knowledge → Projects in the dashboard
2. Click Create Project
3. Enter a name and description
4. Click Create

Step 2: Upload Documents

1. Select your project
2. Click Upload Files
3. Drag and drop or select files:
   • PDF documents
   • Word documents (.docx)
   • Markdown files (.md)
   • HTML files (.html)
   • Plain text (.txt)
4. Files are automatically processed and indexed

Processing includes:

• Text extraction
• Semantic chunking (500-1000 tokens per chunk)
• Embedding generation (OpenAI text-embedding-3-small)
• Summary generation for each topic

Step 3: Enable Knowledge in Your Application

Knowledge Engine integrates with:

• Unified Chat API via demeterics_tools.knowledge
• AI Chat Widget via Knowledge Project attachment

Project Organization

Documents in a Knowledge Project are organized by topics. Each topic contains:

Example structure:

my-project/
├── product-guide/
│   ├── product-guide_summary.md
│   ├── product-guide_detailed.md
│   └── [vector chunks]
├── pricing-policy/
│   ├── pricing-policy_summary.md
│   ├── pricing-policy_detailed.md
│   └── [vector chunks]
└── faq/
    ├── faq_summary.md
    ├── faq_detailed.md
    └── [vector chunks]

How the LLM Uses Knowledge Tools

When the LLM receives a user question, it decides which tools to call based on the query:

Vector Search (search_knowledge)

For specific questions, the LLM searches for relevant passages:

User: "What is the return policy for electronics?"

LLM calls: search_knowledge(query="return policy electronics")

Returns: Top 5 most relevant chunks with similarity scores

Document Discovery (find_documents)

For broad questions, the LLM first discovers which documents might be relevant:

User: "Tell me about your products"

LLM calls: find_documents(query="products")

Returns: List of topics ranked by relevance
- product-guide (5 matches, avg score 0.89)
- pricing-policy (3 matches, avg score 0.76)

Summary Retrieval (get_summary)

For topic overview, the LLM retrieves the summary:

LLM calls: get_summary(topic="product-guide")

Returns: The product-guide_summary.md content

Full Content (get_content)

For detailed information, the LLM retrieves full documents:

LLM calls: get_content(topic="pricing-policy")

Returns: The pricing-policy_detailed.md content
(automatically split if over 30,000 tokens)

Vector Search Details

The Knowledge Engine uses:

• Embedding Model: OpenAI text-embedding-3-small (1536 dimensions)
• Similarity Metric: Cosine similarity
• Chunk Size: 500-1000 tokens with semantic boundaries
• Storage: Google BigQuery with ML.VECTOR_SEARCH

Search flow:

1. Query embedded using same model
2. Cosine similarity computed against all chunks
3. Top-K results returned (default: 5)
4. Results include: content, similarity score, topic, token count

Best Practices

Document Preparation

1. Use clear headings: Structure documents with H1, H2, H3 headers
2. One topic per file: Don't mix unrelated content in a single document
3. Include metadata: Add title, date, and version information
4. Remove boilerplate: Strip headers/footers that repeat across pages
5. Update regularly: Re-upload documents when content changes

Project Organization

1. Separate by domain: Create different projects for different use cases
2. Use descriptive names: customer-support-kb not project-1
3. Version control: Keep source documents in version control
4. Monitor quality: Review LLM responses and refine documents

Query Optimization

1. System prompts: Instruct the LLM when to use knowledge tools
2. Temperature: Use lower temperature (0.3-0.5) for factual retrieval
3. Max iterations: Allow 2-3 tool calls for complex questions
4. Fallback behavior: Define what to do when no relevant content found

Integration Options

Option 1: Unified Chat API (Recommended)

Use the /chat endpoint with demeterics_tools.knowledge enabled.

See: Unified Chat API with Knowledge

Option 2: AI Chat Widget

Attach a Knowledge Project to your AI Chat Widget.

See: AI Chat Widget Knowledge Integration

Option 3: Direct API Access

Query your knowledge base directly via the Knowledge API.

Search endpoint:

curl -X POST https://api.demeterics.com/v1/projects/{project_id}/agent/search \
  -H "Authorization: Bearer dmt_your_api_key" \
  -H "Content-Type: application/json" \
  -d '{"query": "return policy", "top_k": 5}'

Response:

{
  "results": [
    {
      "content": "Our return policy allows returns within 30 days...",
      "topic_slug": "policies",
      "similarity": 0.92,
      "tokens": 156
    }
  ],
  "query_tokens": 3,
  "search_time_ms": 45
}

Pricing

Knowledge Engine usage is billed via Demeterics credits:

All LLM calls made during tool-calling loops are billed at standard rates for the model used.

Limits

Contact sales@demeterics.com for enterprise limits.

Troubleshooting

Documents not appearing in search?

• Check the project's indexing status in the dashboard
• Re-sync vectors: Project → Settings → Resync Vectors
• Verify the document was processed without errors

Low relevance scores?

• Use more descriptive queries
• Check document structure (headings, paragraphs)
• Consider splitting large documents into focused topics

LLM not calling tools?

• Add system prompt: "Use knowledge tools to answer questions about our products"
• Check tool_choice is set to "auto" not "none"
• Verify demeterics_tools.knowledge is enabled

Need help?

• Email: support@demeterics.com
• Dashboard chat widget
• API Reference

Related Documentation

• Unified Chat API – Using Knowledge with the Chat API
• AI Chat Widget – Widget integration with Knowledge
• API Reference – Full API documentation
• Getting Started – Quick start guide