AI-Powered Search¶

Search your emails by meaning, not just keywords. "Find that message about the quarterly budget" actually works.

The RAG (Retrieval-Augmented Generation) service (port 8090) turns email content into vector embeddings and stores them in a Qdrant vector database. When you search, your query is also embedded and compared against stored vectors using cosine similarity. The result: semantic search that understands what you mean, even when you don't use the exact words from the email.

How it works¶

flowchart LR
    subgraph Indexing
        A[New Email] --> B[Chunk text]
        B --> C[Generate embeddings]
        C --> D[Store in Qdrant]
    end

    subgraph Searching
        E[Search query] --> F[Embed query]
        F --> G[Cosine similarity\nsearch in Qdrant]
        G --> H[Return ranked\nresults]
    end

    D -.->|same collection| G

The pipeline, step by step¶

1. Chunking. Long emails are split into chunks (default: 1,000 characters with 200-character overlap). This is important because embedding models have sequence length limits, and smaller chunks give more precise search results.

2. Embedding. Each chunk is converted into a dense vector (a list of numbers) by an embedding model. Mailyte supports multiple providers -- local models run on your hardware with zero data leaving your server, or you can use cloud APIs for higher quality.

3. Storage. Vectors are stored in Qdrant, a purpose-built vector database. Each organization gets its own Qdrant collection for data isolation.

4. Search. Your search query goes through the same embedding model, producing a query vector. Qdrant finds the stored vectors most similar to the query vector (cosine distance) and returns the matching email chunks with relevance scores.

Embedding providers¶

The default (sentence_transformers) runs entirely on your server. Your email content never leaves your infrastructure.

Configuration¶

Core settings¶

Qdrant settings¶

Embedding settings¶

Chunking settings¶

Search settings¶

What gets indexed¶

API endpoints¶

The RAG service is a FastAPI application, so it comes with interactive docs at http://localhost:8090/docs.

Search emails¶

curl -X POST http://localhost:8090/api/v1/organizations/org_123/search \
  -H "Content-Type: application/json" \
  -d '{
    "query": "quarterly budget review meeting",
    "top_k": 5,
    "score_threshold": 0.7
  }'

{
  "results": [
    {
      "email_id": "msg_789",
      "subject": "Q3 Budget Review - Action Items",
      "from": "finance@company.com",
      "score": 0.92,
      "chunk": "...the quarterly budget review highlighted several areas where we need to cut spending...",
      "metadata": {
        "date": "2026-03-10",
        "has_attachments": true
      }
    }
  ],
  "total": 1,
  "search_time_ms": 45
}

Index emails¶

curl -X POST http://localhost:8090/api/v1/organizations/org_123/index \
  -H "Content-Type: application/json" \
  -d '{
    "emails": [
      {
        "email_id": "msg_789",
        "subject": "Q3 Budget Review",
        "from": "finance@company.com",
        "body": "The quarterly budget review highlighted..."
      }
    ]
  }'

Service health¶

curl http://localhost:8090/health/

Returns health status of Qdrant, the embedding model, and overall service readiness.

Admin endpoints¶

# View configuration
curl http://localhost:8090/config

# Admin dashboard
curl http://localhost:8090/admin/stats

Things to know¶

• Vector size must match your embedding model. If you use all-MiniLM-L6-v2, the vectors are 384 dimensions. If you use OpenAI's text-embedding-ada-002, they're 1,536 dimensions. Set QDRANT_VECTOR_SIZE accordingly, or you'll get errors on indexing.

• Privacy mode is on by default. With EMBEDDING_PRIVACY_MODE=true, the service will refuse to use cloud-based embedding providers (OpenAI, Azure, Cohere). This is a safety rail for organizations that can't send email content to third parties. Turn it off explicitly if you want to use cloud embeddings.

• Re-indexing is needed if you change the embedding model. Vectors from different models aren't compatible. If you switch from all-MiniLM-L6-v2 to OpenAI, you need to re-index all emails in the new model's vector space.

• Multi-tenancy means separate Qdrant collections. Each organization gets its own collection (e.g., mailrag_org_123). Search queries are scoped to the requesting org's collection, so Org A can never accidentally find Org B's emails.

• The fallback mechanism helps with resilience. If your primary embedding provider fails (model OOM, API down), the service can fall back to a secondary provider. This means search might use a different model temporarily, and result quality might differ, but the service stays up.

• Qdrant needs its own resources. On a server with millions of emails, the Qdrant database can consume significant memory (it keeps vector indices in RAM for fast search). Plan your server sizing accordingly. A rough guide: 1 million 384-dimensional vectors ~ 1.5 GB RAM.

• Search quality improves with good chunking. The default settings work well for typical emails. But if you're indexing very long emails or attachments, you might want to increase CHUNK_SIZE and CHUNK_OVERLAP for better context preservation. Experiment and see what gives the best results for your data.