AI-Powered Semantic Search

TyneBase's RAG (Retrieval-Augmented Generation) system lets you ask questions and get accurate answers with cited sources.

How RAG Works

User Question
      │
      ▼
┌─────────────────┐
│  Query Embedding │  ← Convert question to vector
└────────┬────────┘
         │
         ▼
┌─────────────────┐
│  Vector Search   │  ← Find similar document chunks
└────────┬────────┘
         │
         ▼
┌─────────────────┐
│  Reranking       │  ← Cross-encoder precision ranking
└────────┬────────┘
         │
         ▼
┌─────────────────┐
│  LLM Generation  │  ← Generate answer with context
└────────┬────────┘
         │
         ▼
   Answer + Sources

Using AI Search

Quick Search

Press ⌘ + K and type your question:

"How do I set up SSO with Okta?"

The AI will:

1. Search your entire knowledge base
2. Find relevant documentation
3. Generate a comprehensive answer
4. Cite the source documents

Advanced Queries

AI search understands context and nuance:

Query Type	Example
How-to	"How do I configure webhooks?"
Comparison	"What's the difference between roles?"
Troubleshooting	"Why is my API returning 403?"
Policy	"What's our data retention policy?"

Search Results

Each result includes:

• Answer: AI-generated response
• Sources: Linked documents used
• Confidence: How certain the AI is
• Related: Suggested follow-up questions

Indexing Configuration

What Gets Indexed

• All published documents
• Document content and metadata
• Code blocks and tables
• Comments (optional)

Chunking Strategy

TyneBase uses semantic chunking:

{
  chunkSize: 512,      // tokens per chunk
  chunkOverlap: 50,    // overlap between chunks
  splitOn: 'semantic', // respect document structure
  preserveTables: true // keep tables intact
}

Hybrid Search

We combine vector search with keyword matching:

Method	Weight	Purpose
Vector	70%	Semantic understanding
Keyword	30%	Exact matches, names

Reranking

After initial retrieval, we use cross-encoder reranking for precision:

1. Retrieve top 50 candidates (fast, broad)
2. Rerank with Cohere Rerank v3.5 (slow, precise)
3. Return top 5 most relevant chunks

Privacy & Compliance

• All AI processing uses EU data centers
• Embeddings are tenant-isolated
• No data leaves your workspace
• Full audit trail of searches