# Embeddings and reranking API

> Vectors for semantic search in Indian languages and English, and a reranker to order results.

For retrieval (RAG) and search without an API key: turn text into 384 numbers with multilingual-e5-small, so texts that mean the same land close together across about a hundred languages, Hindi, Tamil and the other Indian languages among them; then let bge-reranker-v2-m3 read the query with each candidate and order them. Nothing you send is kept or logged.

Base URL `https://api.dharun.dev`. No API key. Answers are JSON, `{ "data": ..., "meta": { ... } }`, and errors are `{ "error": { "code", "message" } }` with a matching HTTP status. Rate limit: 600 requests a minute per client.

## Text embeddings

`GET /v1/embed`, also `POST`

Up to 64 texts a request (GET: one or more text parameters; POST: JSON {"texts": [...], "type": ...}). Each comes back as 384 numbers scaled to length 1, so a dot product is the cosine similarity. Embed questions with type=query and the documents searched with type=passage, as the model was trained. Each text is read up to 512 tokens.

POST body: JSON: {"texts": ["..."], "type": "query" or "passage"}, or {"text": "..."}. Send it as `text/plain`, or as JSON `{"text": "..."}` with the other options beside it.

What you send is never written to the log, and the answer is not cached.

Parameters:

- `text` (query): A text; repeat for several (or POST them).
- `type` (query): query (default) for questions, passage for the documents searched. One of: query, passage.

Example: https://api.dharun.dev/v1/embed?type=query&text=How+do+I+renew+my+passport%3F

Response, with lists and long text shortened:

```json
{
  "data": {
    "model": "multilingual-e5-small",
    "type": "query",
    "dimensions": 384,
    "embeddings": [
      [
        0.056168,
        -0.053384,
        -0.02612
      ]
    ],
    "tokens": [
      14
    ],
    "truncated": []
  },
  "meta": {
    "source": {
      "name": "multilingual-e5-small (int8)",
      "url": "https://huggingface.co/intfloat/multilingual-e5-small",
      "licence": "MIT",
      "version": "multilingual-e5-small@614241f",
      "retrieved": "2026-10-15T12:28:44Z",
      "attribution": "Embeddings by multilingual-e5-small (Wang et al., Multilingual E5 Text Embeddings, 2024), MIT."
    },
    "note": "Each embedding is normalised to length 1, so the dot product of two is their cosine similarity. Embed questions with type=query and the documents searched with …"
  }
}
```

## Rerank documents for a query

`GET /v1/rerank`, also `POST`

Scores how well each document answers the query, from 0 to 1, and returns them best first with their place in the request. Up to 100 documents of up to 8,000 characters; each pair is read up to 512 tokens. GET takes query and repeated document parameters; POST takes JSON.

POST body: JSON: {"query": "...", "documents": ["...", "..."], "top": 3}. Send it as `text/plain`, or as JSON `{"text": "..."}` with the other options beside it.

What you send is never written to the log, and the answer is not cached.

Parameters:

- `query` (query, required): The question, up to 2,000 characters.
- `document` (query): A candidate; repeat for each (or POST them).
- `top` (query): Return only the best few. Default all.

Example: https://api.dharun.dev/v1/rerank?query=capital+of+Tamil+Nadu&document=Chennai+is+the+capital+of+Tamil+Nadu.&document=Mumbai+is+the+financial+capital+of+India.

## Data source

- [multilingual-e5-small (int8)](https://huggingface.co/intfloat/multilingual-e5-small), multilingual-e5-small@614241f, retrieved 2026-10-15T12:28:44Z. Licence: MIT. Embeddings by multilingual-e5-small (Wang et al., Multilingual E5 Text Embeddings, 2024), MIT.