Rerank API | Tutujin API Docs

POST /v1/rerank takes a query and a list of candidate documents, gives each document a relevance score and returns them sorted from most to least relevant. In knowledge-base RAG you typically retrieve a few dozen passages with vector search, then rerank them and pass only the best few to the chat model: answers get more accurate and the chat model reads less irrelevant text. The request uses the common query, documents and top_n fields and is relayed to the model service as-is.

POSThttps://open.tutujin.com/v1/rerank

Headers

Header	Required	Description
`Authorization`	Yes	`Bearer sk-...`, an API key from the console
`Content-Type`	Yes	`application/json`

Request body

Parameter	Type	Required	Description
`model`	string	Yes	Model name, e.g. `qwen3-reranker-8b`
`query`	string	Yes	The query, usually the user's question
`documents`	array	Yes	Candidate documents, as an array of strings
`top_n`	integer	No	Return only the N highest-scoring documents; omit to return all
`return_documents`	boolean	No	`true` includes each document's text in the results

Request example

curl https://open.tutujin.com/v1/rerank \
  -H "Authorization: Bearer $TUTUJIN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen3-reranker-8b",
    "query": "Can I return an item after opening it?",
    "documents": [
      "Unopened items can be returned within 7 days of delivery, no reason needed.",
      "Orders ship within 48 hours of payment; holidays push this back.",
      "Opened items without a quality issue cannot be returned without a reason."
    ],
    "top_n": 2
  }'

import os
import requests

docs = [
    "Unopened items can be returned within 7 days of delivery, no reason needed.",
    "Orders ship within 48 hours of payment; holidays push this back.",
    "Opened items without a quality issue cannot be returned without a reason.",
]
r = requests.post(
    "https://open.tutujin.com/v1/rerank",
    headers={"Authorization": f"Bearer {os.environ['TUTUJIN_API_KEY']}"},
    json={
        "model": "qwen3-reranker-8b",
        "query": "Can I return an item after opening it?",
        "documents": docs,
        "top_n": 2,
    },
    timeout=30,
)
r.raise_for_status()
for item in r.json()["results"]:
    print(round(item["relevance_score"], 3), docs[item["index"]])

const docs = [
  'Unopened items can be returned within 7 days of delivery, no reason needed.',
  'Orders ship within 48 hours of payment; holidays push this back.',
  'Opened items without a quality issue cannot be returned without a reason.',
]
const r = await fetch('https://open.tutujin.com/v1/rerank', {
  method: 'POST',
  headers: {
    Authorization: `Bearer ${process.env.TUTUJIN_API_KEY}`,
    'Content-Type': 'application/json',
  },
  body: JSON.stringify({
    model: 'qwen3-reranker-8b',
    query: 'Can I return an item after opening it?',
    documents: docs,
    top_n: 2,
  }),
})
if (!r.ok) throw new Error(await r.text())
const { results } = await r.json()
for (const item of results) console.log(item.relevance_score.toFixed(3), docs[item.index])

Response example

JSON

{
  "results": [
    { "index": 2, "relevance_score": 0.93 },
    { "index": 0, "relevance_score": 0.71 }
  ],
  "usage": { "prompt_tokens": 96, "total_tokens": 96 }
}

Field	Description
`results[].index`	Position of the document in `documents`, starting at 0
`results[].relevance_score`	Relevance score; higher is more relevant. Ranges differ by model, so compare scores within one request only
`results[].document`	Present when `return_documents` is `true`; contains the document text
`usage`	Billed usage; some models bill by tokens, some per request

Notes

This endpoint is relayed as-is and has not been verified separately on this site; check the response shape with a little traffic first.
Results are already sorted by score; use index to look up the original text.
More and longer candidates mean more usage. Retrieving 20 to 50 passages and keeping a top_n of 3 to 8 is usually enough.
Tune any score threshold on your own data. Scores mean different things on different models, so re-tune after switching models.
The full flow with vector search is in Enterprise knowledge base and RAG.

Models that support this endpoint

Model	Vendor	Cheapest group	Input ¥/1M	Output ¥/1M	Per request ¥
qwen3-reranker-4b	Qwen	Standard	—	—	¥0.002
qwen3-reranker-0.6b	Qwen	Standard	—	—	¥0.002
qwen3-reranker-8b	Qwen	China Models	¥0.00308	¥0.00308	—
jina-reranker-v2-base-multilingual	Other	Standard	¥0.4	¥0.4	—
netease-youdao/bce-reranker-base_v1	Other	Standard	¥0.4	¥0.4	—
BAAI/bge-reranker-v2-m3	Other	Standard	¥0.4	¥0.4	—

Prices in CNY, cheapest group for each model; updated Nov 12, 2026, 8:17 PM.

Still stuck? Add our WeChat support JsnonoChat and send the request id from the error message, and we will look into it.

Header

Required

Description

Authorization

Yes

Bearer sk-..., an API key from the console

Content-Type

Yes

application/json

Request body

Parameter	Type	Required	Description
`model`	string	Yes	Model name, e.g. `qwen3-reranker-8b`
`query`	string	Yes	The query, usually the user's question
`documents`	array	Yes	Candidate documents, as an array of strings
`top_n`	integer	No	Return only the N highest-scoring documents; omit to return all
`return_documents`	boolean	No	`true` includes each document's text in the results

Request example

curl https://open.tutujin.com/v1/rerank \
  -H "Authorization: Bearer $TUTUJIN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen3-reranker-8b",
    "query": "Can I return an item after opening it?",
    "documents": [
      "Unopened items can be returned within 7 days of delivery, no reason needed.",
      "Orders ship within 48 hours of payment; holidays push this back.",
      "Opened items without a quality issue cannot be returned without a reason."
    ],
    "top_n": 2
  }'

import os
import requests

docs = [
    "Unopened items can be returned within 7 days of delivery, no reason needed.",
    "Orders ship within 48 hours of payment; holidays push this back.",
    "Opened items without a quality issue cannot be returned without a reason.",
]
r = requests.post(
    "https://open.tutujin.com/v1/rerank",
    headers={"Authorization": f"Bearer {os.environ['TUTUJIN_API_KEY']}"},
    json={
        "model": "qwen3-reranker-8b",
        "query": "Can I return an item after opening it?",
        "documents": docs,
        "top_n": 2,
    },
    timeout=30,
)
r.raise_for_status()
for item in r.json()["results"]:
    print(round(item["relevance_score"], 3), docs[item["index"]])

const docs = [
  'Unopened items can be returned within 7 days of delivery, no reason needed.',
  'Orders ship within 48 hours of payment; holidays push this back.',
  'Opened items without a quality issue cannot be returned without a reason.',
]
const r = await fetch('https://open.tutujin.com/v1/rerank', {
  method: 'POST',
  headers: {
    Authorization: `Bearer ${process.env.TUTUJIN_API_KEY}`,
    'Content-Type': 'application/json',
  },
  body: JSON.stringify({
    model: 'qwen3-reranker-8b',
    query: 'Can I return an item after opening it?',
    documents: docs,
    top_n: 2,
  }),
})
if (!r.ok) throw new Error(await r.text())
const { results } = await r.json()
for (const item of results) console.log(item.relevance_score.toFixed(3), docs[item.index])

Response example

JSON

{
  "results": [
    { "index": 2, "relevance_score": 0.93 },
    { "index": 0, "relevance_score": 0.71 }
  ],
  "usage": { "prompt_tokens": 96, "total_tokens": 96 }
}

Field	Description
`results[].index`	Position of the document in `documents`, starting at 0
`results[].relevance_score`	Relevance score; higher is more relevant. Ranges differ by model, so compare scores within one request only
`results[].document`	Present when `return_documents` is `true`; contains the document text
`usage`	Billed usage; some models bill by tokens, some per request

Notes

This endpoint is relayed as-is and has not been verified separately on this site; check the response shape with a little traffic first.

Results are already sorted by score; use index to look up the original text.

More and longer candidates mean more usage. Retrieving 20 to 50 passages and keeping a top_n of 3 to 8 is usually enough.

Tune any score threshold on your own data. Scores mean different things on different models, so re-tune after switching models.

The full flow with vector search is in Enterprise knowledge base and RAG.

Models that support this endpoint

Model	Vendor	Cheapest group	Input ¥/1M	Output ¥/1M	Per request ¥
qwen3-reranker-4b	Qwen	Standard	—	—	¥0.002
qwen3-reranker-0.6b	Qwen	Standard	—	—	¥0.002
qwen3-reranker-8b	Qwen	China Models	¥0.00308	¥0.00308	—
jina-reranker-v2-base-multilingual	Other	Standard	¥0.4	¥0.4	—
netease-youdao/bce-reranker-base_v1	Other	Standard	¥0.4	¥0.4	—
BAAI/bge-reranker-v2-m3	Other	Standard	¥0.4	¥0.4	—

Prices in CNY, cheapest group for each model; updated Nov 12, 2026, 8:17 PM.