Overview

The Vectara RAG-as-a-Service MCP (Model Context Protocol) server is a lightweight bridge that connects your applications to Vectara’s retrieval-augmented generation pipeline. It exposes a secure, production-ready endpoint that accepts query requests, retrieves relevant context from indexed corpora, and returns structured candidate passages or augmented answers suitable for use with an LLM. The MCP server is designed for predictable latency, multi-tenant deployments, and enterprise-grade access controls.

This server is useful when you want to separate retrieval and context management from LLM inference: it standardizes context fetch and scoring, centralizes access policies, and simplifies scaling retrieval independently of model serving. Developers use MCP to get fast, relevant evidence for answering questions, to build automated support agents, and to power internal knowledge search with auditable request/response semantics.

Features

Retrieval-augmented generation (RAG) gateway that returns scored context passages and citations
Secure API with API-key or bearer token support and configurable auth backends
Multi-tenant and multi-corpus support for enterprise deployments
Metrics and health endpoints for observability and integration with monitoring systems
Docker-friendly and configurable via environment variables or a YAML config file
Simple JSON API that integrates with any LLM or downstream orchestration layer

Installation / Configuration

Clone the repository and run using Docker (recommended) or a prebuilt binary.

Clone the repo

git clone https://github.com/vectara/vectara-mcp.git
cd vectara-mcp

Build and run with Docker

# Build (if you want to create a local image)
docker build -t vectara-mcp:latest .

# Run with environment variables from a file
docker run --env-file .env -p 8080:8080 vectara-mcp:latest

Run with a prebuilt image from a container registry (example)

docker run --env-file .env -p 8080:8080 ghcr.io/vectara/vectara-mcp:latest

Example .env configuration

# Vectara account and corpus identifiers
VECTARA_CUSTOMER_ID=123456789
VECTARA_CORPUS_ID=example-corpus

# API credentials for Vectara
VECTARA_API_KEY=sk_live_...

# MCP server settings
MCP_PORT=8080
MCP_HOST=0.0.0.0

# Optional TLS and auth
TLS_CERT=/etc/ssl/certs/mcp.crt
TLS_KEY=/etc/ssl/private/mcp.key
API_KEYS=client1:secret1,client2:secret2

Example YAML config (alternative to env vars)

server:
  host: 0.0.0.0
  port: 8080
vectara:
  customerId: 123456789
  corpusId: example-corpus
  apiKey: sk_live_...
auth:
  type: api_keys
  keys:
    - id: client1
      secret: secret1

Available Resources

The MCP server exposes a small set of HTTP endpoints for integration:

Endpoint	Method	Purpose
/mcp/query	POST	Run a retrieval request; returns contexts, scores, and metadata
/mcp/health	GET	Liveness/Readiness check for orchestration
/mcp/metrics	GET	Prometheus-compatible metrics (if enabled)
/mcp/config	GET (admin)	Return effective configuration (restricted)

Example query request (JSON)

curl -X POST "http://localhost:8080/mcp/query" \
  -H "Authorization: Bearer client-secret" \
  -H "Content-Type: application/json" \
  -d '{
    "query": "How do I reset my account password?",
    "top_k": 5,
    "filters": {"language": "en"}
  }'

Example response (truncated)

{
  "query": "How do I reset my account password?",
  "results": [
    {
      "id": "doc-123",
      "score": 0.98,
      "text": "To reset your password, go to Settings > Security...",
      "source": "knowledge-base/article-42"
    }
  ],
  "meta": {
    "took_ms": 35
  }
}

SDKs and integrations

Use any HTTP-capable SDK (curl, requests, fetch). You can wrap the MCP output and feed it to your LLM prompt builder.
Check the GitHub repo for example clients and CI/CD deployment manifests: https://github.com/vectara/vectara-mcp

Use Cases

Customer support agent: Query internal FAQs and return top passages to an LLM to produce concise, cited answers. Example flow: user question → MCP query → LLM synthesize with citations → respond to user.
Secure internal search: Only allow requests from authorized clients with API keys, and funnel all retrieval through MCP for auditing and rate-limiting.
Multi-tenant knowledge access: Host corpora for multiple business units and route requests to the appropriate corpus using configuration or request-scoped metadata.
Real-time operations: Use MCP’s low-latency retrieval to enrich live chat or voice assistants with accurate, up-to-date documentation excerpts.

Tips for Developers

Start by indexing the documents you want Vectara to search; MCP expects corpora to be available in your Vectara account.
Tune top_k and filters to balance recall vs. latency for downstream LLM prompts.
Use /mcp/health and /mcp/metrics in your orchestration to provide automated restarts and capacity planning.
Protect admin endpoints and configuration access; rotate your Vectara API key periodically.

For full implementation details, examples, and advanced deployment patterns, see the repository: https://github.com/vectara/vectara-mcp.

Vectara RAG-as-a-Service MCP Server