Embed and rerank

Conare’s embedding and rerank models also work without namespaces, through OpenAI- and Cohere-compatible endpoints. Point your existing client at Conare. A service that only embeds and reranks can use a Models only key, which reaches no namespace.

Embed

POST /v1/embeddings

curl https://api.conare.ai/v1/embeddings \
  -H "Authorization: Bearer $CONARE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "e5-mine",
    "input": ["Refunds are available for 30 days."]
  }'

import OpenAI from "openai"; // npm install openai

const client = new OpenAI({
  baseURL: "https://api.conare.ai/v1",
  apiKey: process.env.CONARE_API_KEY,
});

const docs = await client.embeddings.create({
  model: "e5-mine",
  input: ["Refunds are available for 30 days."],
});
const query = await client.embeddings.create({
  model: "e5-mine-query",
  input: "how do refunds work?",
});

import os
from openai import OpenAI  # pip install openai

client = OpenAI(
    base_url="https://api.conare.ai/v1",
    api_key=os.environ["CONARE_API_KEY"],
)

docs = client.embeddings.create(
    model="e5-mine",
    input=["Refunds are available for 30 days."],
)
query = client.embeddings.create(
    model="e5-mine-query",
    input="how do refunds work?",
)

Response

{
  "object": "list",
  "data": [
    {
      "object": "embedding",
      "index": 0,
      "embedding": [0.0129, 0.0584, 0.0144]
    }
  ],
  "model": "e5-mine",
  "usage": { "prompt_tokens": 11, "total_tokens": 11 }
}

Request

input (string or array, required): 1-128 texts.
model (string, default e5-mine): e5-mine for text you store, e5-mine-query for searches.
encoding_format (string, default float): float or base64.

Response

data: One embedding per input, in order. Each has 1,024 numbers (shortened above).
usage: The tokens read.

Only the first 6,000 characters of each text are embedded. A query’s vector includes today’s date, so it changes from day to day. LangChain’s OpenAIEmbeddings needs check_embedding_ctx_length=False, or it sends token ids instead of text.

Rerank

POST /v1/rerank

curl https://api.conare.ai/v1/rerank \
  -H "Authorization: Bearer $CONARE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "query": "how do refunds work?",
    "documents": [
      "Orders ship within two days.",
      "Refunds are available for 30 days."
    ],
    "top_n": 1
  }'

const response = await fetch("https://api.conare.ai/v1/rerank", {
  method: "POST",
  headers: {
    Authorization: `Bearer ${process.env.CONARE_API_KEY}`,
    "Content-Type": "application/json",
  },
  body: JSON.stringify({
    query: "how do refunds work?",
    documents: [
      "Orders ship within two days.",
      "Refunds are available for 30 days.",
    ],
    top_n: 1,
  }),
});

import os
import cohere  # pip install cohere

co = cohere.Client(
    base_url="https://api.conare.ai",
    api_key=os.environ["CONARE_API_KEY"],
)

response = co.rerank(
    query="how do refunds work?",
    documents=[
        "Orders ship within two days.",
        "Refunds are available for 30 days.",
    ],
    top_n=1,
)

Response

{
  "id": "81bd9721-2c65-43b9-97cc-dd3141baf768",
  "model": "conare-rerank-gen2-...",
  "results": [{ "index": 1, "relevance_score": 0.38 }],
  "unscored": [],
  "usage": { "total_tokens": 57 }
}

Request

query (string, required): Up to 8,192 characters.
documents (array, required): 1-100 documents, as strings or { "text": ... } objects. Each is up to 24,000 characters, and together with the query up to 256,000.
top_n (integer, default all): Return only the best top_n.
return_documents (boolean, default false): Add each document’s text to its result.

Response

results: Best first. index is the document’s position in documents, and relevance_score is higher for a better match.
unscored: Documents the reranker didn’t reach in time. They aren’t in results.
usage: The tokens read.

Use Cohere’s v1 client, cohere.Client. cohere.ClientV2 calls /v2/rerank, which Conare doesn’t serve.

Models

GET /v1/models lists the models your key can call.

Model	Use
`e5-mine`	Embed text you store
`e5-mine-query`	Embed searches
`conare-rerank-...`	Rerank. Leave `model` out to use it.

Models trained for your organization are listed too.

With the Conare SDK

const [vector] = await conare.embed("how do refunds work?", {
  type: "query",
});

const ranked = await conare.rerank("how do refunds work?", [
  "Orders ship within two days.",
  "Refunds are available for 30 days.",
]);
// [{ index: 1, score: 0.38, document: "Refunds are..." }, ...]

These have the same limits as the /v1 routes.

Get started

API

Tools

Embed and rerank

Embed

Request

Response

Rerank

Request

Response

Models

With the Conare SDK

​Embed

​Request

​Response

​Rerank

​Request

​Response

​Models

​With the Conare SDK

Embed

Request

Response

Rerank

Request

Response

Models

With the Conare SDK