执行推理 API
编辑执行推理 API编辑
此功能处于技术预览阶段，可能会在未来版本中更改或删除。Elastic 将努力解决任何问题，但技术预览版中的功能不受官方 GA 功能支持 SLA 的约束。
使用推理端点对输入文本执行推理任务。
推理 API 使您能够使用某些服务，例如内置机器学习模型（ELSER、E5）、通过 Eland、Cohere、OpenAI 或 Hugging Face 上传的模型。对于内置模型和通过 Eland 上传的模型，推理 API 提供了一种使用和管理训练模型的替代方法。但是，如果您不打算使用推理 API 来使用这些模型，或者您想使用非 NLP 模型，请使用机器学习训练模型 API。
请求编辑
POST /_inference/<inference_id>
POST /_inference/<task_type>/<inference_id>
先决条件编辑
- 需要
monitor_inference集群权限（内置的inference_admin和inference_user角色授予此权限）
说明编辑
执行推理 API 使您能够使用机器学习模型对您作为输入提供的数据执行特定任务。API 返回包含任务结果的响应。您使用的推理端点可以执行一项特定任务，该任务是在使用创建推理 API创建端点时定义的。
路径参数编辑
-
<inference_id> - （必填，字符串）推理端点的唯一标识符。
-
<task_type> - （可选，字符串）模型执行的推理任务类型。
查询参数编辑
-
timeout - （可选，超时）控制等待推理完成的时间量。默认为 30 秒。
请求正文编辑
-
input -
（必填，字符串或字符串数组）要对其执行推理任务的文本。
input可以是单个字符串或数组。completion任务类型的推理端点当前仅支持单个字符串作为输入。 -
query - （必填，字符串）仅适用于
rerank推理端点。搜索查询文本。
示例编辑
完成示例编辑
以下示例对示例问题执行完成。
resp = client.inference.inference(
task_type="completion",
inference_id="openai_chat_completions",
body={"input": "What is Elastic?"},
)
print(resp)
POST _inference/completion/openai_chat_completions
{
"input": "What is Elastic?"
}
API 返回以下响应
{
"completion": [
{
"result": "Elastic is a company that provides a range of software solutions for search, logging, security, and analytics. Their flagship product is Elasticsearch, an open-source, distributed search engine that allows users to search, analyze, and visualize large volumes of data in real-time. Elastic also offers products such as Kibana, a data visualization tool, and Logstash, a log management and pipeline tool, as well as various other tools and solutions for data analysis and management."
}
]
}
重新排序示例编辑
以下示例对示例输入执行重新排序。
resp = client.inference.inference(
task_type="rerank",
inference_id="cohere_rerank",
body={
"input": ["luke", "like", "leia", "chewy", "r2d2", "star", "wars"],
"query": "star wars main character",
},
)
print(resp)
POST _inference/rerank/cohere_rerank
{
"input": ["luke", "like", "leia", "chewy","r2d2", "star", "wars"],
"query": "star wars main character"
}
API 返回以下响应
{
"rerank": [
{
"index": "2",
"relevance_score": "0.011597361",
"text": "leia"
},
{
"index": "0",
"relevance_score": "0.006338922",
"text": "luke"
},
{
"index": "5",
"relevance_score": "0.0016166499",
"text": "star"
},
{
"index": "4",
"relevance_score": "0.0011695103",
"text": "r2d2"
},
{
"index": "1",
"relevance_score": "5.614787E-4",
"text": "like"
},
{
"index": "6",
"relevance_score": "3.7850367E-4",
"text": "wars"
},
{
"index": "3",
"relevance_score": "1.2508839E-5",
"text": "chewy"
}
]
}
稀疏嵌入示例编辑
以下示例对示例句子执行稀疏嵌入。
resp = client.inference.inference(
task_type="sparse_embedding",
inference_id="my-elser-model",
body={
"input": "The sky above the port was the color of television tuned to a dead channel."
},
)
print(resp)
response = client.inference.inference(
task_type: 'sparse_embedding',
inference_id: 'my-elser-model',
body: {
input: 'The sky above the port was the color of television tuned to a dead channel.'
}
)
puts response
POST _inference/sparse_embedding/my-elser-model
{
"input": "The sky above the port was the color of television tuned to a dead channel."
}
API 返回以下响应
{
"sparse_embedding": [
{
"port": 2.1259406,
"sky": 1.7073475,
"color": 1.6922266,
"dead": 1.6247464,
"television": 1.3525393,
"above": 1.2425821,
"tuned": 1.1440028,
"colors": 1.1218185,
"tv": 1.0111054,
"ports": 1.0067928,
"poem": 1.0042328,
"channel": 0.99471164,
"tune": 0.96235967,
"scene": 0.9020516,
(...)
},
(...)
]
}