Skip to content

RAG Service

Introduction

Devant provides a set of RESTful API endpoints for Retrieval-Augmented Generation (RAG) workflows. These endpoints enable you to ingest, retrieve, and process documents programmatically.

Step 1: Create service

Navigate to your organization using the Organization dropdown in the top left of the Devant console header. In the left navigation menu, click RAG, then select Service.

Fill in the required fields in the Create RAG Ingestion Service form to set up a new RAG service in Devant.

Field Value
Project Select the target project from dropdown which lists the available projects in your organization
Display Name Sample RAG Service
Name sample-rag-service
Description (Optional) My rag service description

Click the Create Service button and your service will get created

Note

  • When the service is created, Devant automatically increases the container resources (CPU and memory) to ensure reliable operation.

Step 2: Test endpoints

Once the component is created you will be redirected to the Overview page.

  1. On the development environment card, click Test to open the OpenAPI Console, where you will be able to try out all the available endpoints from the endpoint list.
  2. Expand the resource you want to test.
  3. Click Try it out to enable it.
  4. Provide values for the parameters.
  5. Click Execute. The response will be displayed under the Responses section.

Note

  • Some parameters are automatically populated with default values. You can modify them as needed.

RAG Service

Available API endpoints

The image below shows how all the available endpoints are listed in the Console page. You can expand each endpoint and try them out as needed.

RAG Service

POST /upload

Upload a file and ingest it into your vector store. Supports PDF (including scanned PDFs), DOCX, PPTX, XLSX, CSV, HTML, MD, images, and audio (MP3, WAV, M4A, FLAC, OGG).

Required in the request:

  • File to upload.
  • Vector DB provider (e.g., Pinecone, Chroma, Weaviate, Postgres) and connection/API key details.
  • Collection name: where data will be stored.
  • Embedding model provider and model (e.g., OpenAI, Azure, Mistral) and API key.
  • Chunking strategy: recursive, sentence, or character.
  • Max segment size and max overlap size for chunking.

Expected response:

Returns a JSON object indicating successful ingestion, including the file name and type. 

{
    "message": "Added data to vector store successfully",
    "filename": "example.pdf",
    "file_type": "document"
}

POST /retrieve

Retrieve relevant chunks from your vector store based on a user query. Supports semantic search and optional reranking with Cohere.

Required in the request:

  • Vector DB provider and connection/API key details.
  • Name of the collection from which you want to retrieve chunks.
  • Embedding model provider and model, and API key.
  • User query for which you want to retrieve chunks.
  • Max number of chunks to retrieve and minimum similarity threshold.

Optional:

  • Cohere re-ranking model and API key (if using reranking) and the number of top results to rerank.
Info

Expected response:

Returns a JSON object containing the user query and an array of retrieved chunks. Each chunk includes the content, the file it is from, and a timestamp.

{
    "query": "What is Devant?",
    "retrieved_chunks": [
        {
            "text": "Devant is ...",
            "source": "example.pdf",
            "timestamp": "2026-02-16T12:02:25.076312"
        },
        ...
    ]
}

POST /chunks

Parse and chunk an uploaded file, returning the chunks as a JSON array. Does not store data in the vector DB.

This endpoint accepts all the same file types as the /upload endpoint, including PDF (with scanned PDF support), DOCX, PPTX, XLSX, CSV, HTML, MD, images, and audio files (MP3, WAV, M4A, FLAC, OGG).

Required in the request:

  • File to upload.
  • Chunk type: recursive, sentence, or character.
  • Max chunk size and max overlap size.

Expected response:

Returns a JSON object containing the file name and an array of chunks. Each chunk includes a chunk ID and its content.

{
    "filename": "example.pdf",
    "chunks": [
        { "chunk_id": 0, "content": "First chunk content..." },
        { "chunk_id": 1, "content": "Second chunk content..." }
    ]
}

GET /health

Health check endpoint.

Expected response:

Returns a JSON object indicating the service status.

{
    "status": "ok"
}
Info
  • For more details on RAG ingestion and retrieval, and how to obtain API keys and credentials refer the RAG Ingestion guide.