Monte Carlo RAG

Dexrag uses Monte Carlo Tree Search (MCTS) - the same algorithm behind AlphaGo - to intelligently explore documents and find the most relevant content.

The Embedding Problem

Traditional RAG systems use embeddings to find similar documents. But Google DeepMind research revealed fundamental limitations:

Mathematical Ceiling

• Fixed dimensions: 4096 dimensions can only represent ~250M unique documents
• Semantic collapse: Similar concepts get forced into nearby vectors
• Degrading recall: Performance drops as document count increases
• LIMIT benchmark: GPT-4 embeddings achieve <20% recall on complex queries

Static Training

• Internet-trained: Models trained on Reddit and Wikipedia
• Domain blind: Can't adapt to legal, medical, or technical terminology
• No learning: Performance never improves with usage

How Monte Carlo RAG Works

1. BM25 Baseline

Start with a proven text search algorithm (BM25) that doesn't require embeddings.

2. Monte Carlo Exploration

Instead of finding "similar" documents, MCTS:

• Explores document structure intelligently
• Evaluates multiple paths through the content
• Selects the most promising areas to investigate
• Expands understanding based on what's found

3. Document Structure Understanding

Dexrag analyzes:

• Headings and sections
• Tables and lists
• Citations and references
• Paragraph relationships

4. Adaptive Learning

After each query, the system learns:

• Which documents users found relevant
• Common search patterns
• Domain-specific terminology
• User preferences

Performance Comparison

Why It's Better

Breaks Mathematical Limits

Monte Carlo search has no fixed dimensions - it can scale to billions of documents without degrading performance.

Learns Your Domain

Instead of being stuck with Reddit training data, Dexrag adapts to:

• Your industry terminology
• Your document structure
• Your users' needs

Explainable Results

Every search returns an exploration tree showing:

• Which paths were explored
• Why documents were selected
• How relevance was determined

No more black box similarity scores.

Technical Details

Search Algorithm

function monteCarloSearch(query, documents):
  tree = initializeTree(documents)

  for iteration in maxIterations:
    node = selectPromisingNode(tree)
    node = expandNode(node, query)
    value = evaluateNode(node)
    backpropagate(tree, node, value)

  return bestPath(tree)

Adaptive Learning

The system tracks:

• Click-through rate per document
• Time spent on each result
• Follow-up queries
• Explicit relevance feedback

This data updates the selection probabilities for future searches.

When to Use Monte Carlo RAG

Monte Carlo RAG excels at:

• Large document sets (>100K docs)
• Domain-specific content (legal, medical, technical)
• Structured documents (contracts, reports, papers)
• Repeated queries (customer support, internal wikis)

Traditional embeddings may still work for:

• Very small document sets (<100 docs)
• Generic content (news, blogs)
• One-time searches

But even then, Dexrag performs as well or better from query #1.