Link Search Menu Expand Document Documentation Menu
Documentation

ML inference search response processor

Introduced 2.16

The ml_inference search response processor is used to invoke registered machine learning (ML) models in order to incorporate their outputs as new fields in documents within search results.

PREREQUISITE
Before using the ml_inference search response processor, you must have either a local ML model hosted on your OpenSearch cluster or an externally hosted model connected to your OpenSearch cluster through the ML Commons plugin. For more information about local models, see Using ML models within OpenSearch. For more information about externally hosted models, see Connecting to externally hosted models.

Syntax

The following is the syntax for the ml-inference search response processor:

{
  "ml_inference": {
    "model_id": "<model_id>",
    "function_name": "<function_name>",
    "full_response_path": "<full_response_path>",
    "model_config":{
      "<model_config_field>": "<config_value>"
    },
    "model_input": "<model_input>",
    "input_map": [
      {
        "<model_input_field>": "<document_field>"
      }
    ],
    "output_map": [
      {
        "<new_document_field>": "<model_output_field>"
      }
    ],
    "override": "<override>",
    "one_to_one": false
  }
}

Request body fields

The following table lists the required and optional parameters for the ml-inference search response processor.

Parameter Data type Required/Optional Description
model_id String Required The ID of the ML model used by the processor.
function_name String Optional for externally hosted models

Required for local models		 The function name of the ML model configured in the processor. For local models, valid values are sparse_encoding, sparse_tokenize, text_embedding, and text_similarity. For externally hosted models, valid value is remote. Default is remote.
model_config Object Optional Custom configuration options for the ML model. For more information, see The model_config object.
model_input String Optional for externally hosted models

Required for local models		 A template that defines the input field format expected by the model. Each local model type might use a different set of inputs. For externally hosted models, default is "{ \"parameters\": ${ml_inference.parameters} }.
input_map Array Optional for externally hosted models

Required for local models		 An array specifying how to map document fields in the search response to the model input fields. Each element of the array is a map in the "<model_input_field>": "<document_field>" format and corresponds to one model invocation of a document field. If no input mapping is specified for an externally hosted model, then all document fields are passed to the model directly as input. The input_map size indicates the number of times the model is invoked (the number of Predict API requests).
<model_input_field> String Optional for externally hosted models

Required for local models		 The model input field name.
<document_field> String Optional for externally hosted models

Required for local models		 The name or JSON path of the document field in the search response used as the model input.
output_map Array Optional for externally hosted models

Required for local models		 An array specifying how to map the model output fields to new fields in the search response document. Each element of the array is a map in the "<new_document_field>": "<model_output_field>" format.
<new_document_field> String Optional for externally hosted models

Required for local models		 The name of the new field in the document in which the model’s output (specified by model_output) is stored. If no output mapping is specified for externally hosted models, then all fields from the model output are added to the new document field.
<model_output_field> String Optional for externally hosted models

Required for local models		 The name or JSON path of the field in the model output to be stored in the new_document_field.
full_response_path Boolean Optional Set this parameter to true if the model_output_field contains a full JSON path to the field instead of the field name. The model output will then be fully parsed to get the value of the field. Default is true for local models and false for externally hosted models.
ignore_missing Boolean Optional If true and any of the input fields defined in the input_map or output_map are missing, then the missing fields are ignored. Otherwise, a missing field causes a failure. Default is false.
ignore_failure Boolean Optional Specifies whether the processor continues execution even if it encounters an error. If true, then any failure is ignored and the search continues. If false, then any failure causes the search to be canceled. Default is false.
override Boolean Optional Relevant if a document in the response already contains a field with the name specified in <new_document_field>. If override is false, then the input field is skipped. If true, then the existing field value is overridden by the new model output. Default is false.
max_prediction_tasks Integer Optional The maximum number of concurrent model invocations that can run during document search. Default is 10.
one_to_one Boolean Optional Set this parameter to true to invoke the model once (make one Predict API request) for each document. Default value (false) specifies to invoke the model with all documents from the search response, making one Predict API request.
description String Optional A brief description of the processor.
tag String Optional An identifier tag for the processor. Useful for debugging to distinguish between processors of the same type.

The input_map and output_map mappings support standard JSON path notation for specifying complex data structures.

Setup

Create an index named my_index and index one document to explain the mappings:

POST /my_index/_doc/1
{
  "passage_text": "hello world"
}

Using the processor

Follow these steps to use the processor in a pipeline. You must provide a model ID when creating the processor. Before testing a pipeline using the processor, make sure that the model is successfully deployed. You can check the model state using the Get Model API.

For local models, you must provide a model_input field that specifies the model input format. Add any input fields in model_config to model_input.

For remote models, the model_input field is optional, and its default value is "{ \"parameters\": ${ml_inference.parameters} }.

Example: Local model

The following example shows you how to configure an ml_inference search response processor with a local model.

Step 1: Create a pipeline

The following example shows you how to create a search pipeline for the huggingface/sentence-transformers/all-distilroberta-v1 local model. The model is a pretrained sentence transformer model hosted in your OpenSearch cluster.

If you invoke the model using the Predict API, then the request appears as follows:

POST /_plugins/_ml/_predict/text_embedding/cleMb4kBJ1eYAeTMFFg4
{
  "text_docs":[ "today is sunny"],
  "return_number": true,
  "target_response": ["sentence_embedding"]
}

Using this schema, specify the model_input as follows:

 "model_input": "{ \"text_docs\": ${input_map.text_docs}, \"return_number\": ${model_config.return_number}, \"target_response\": ${model_config.target_response} }"

In the input_map, map the passage_text document field to the text_docs field expected by the model:

"input_map": [
  {
    "text_docs": "passage_text"
  }
]

Because you specified the field to be converted into embeddings as a JSON path, you need to set the full_response_path to true. Then the full JSON document is parsed in order to obtain the input field:

"full_response_path": true

The text in the passage_text field will be used to generate embeddings:

{
  "passage_text": "hello world"
}

The Predict API request returns the following response:

{
  "inference_results" : [
    {
      "output" : [
        {
          "name" : "sentence_embedding",
          "data_type" : "FLOAT32",
          "shape" : [
            768
          ],
          "data" : [
            0.25517133,
            -0.28009856,
            0.48519906,
            ...
          ]
        }
      ]
    }
  ]
}

The model generates embeddings in the $.inference_results.*.output.*.data field. The output_map maps this field to the newly created passage_embedding field in the search response document:

"output_map": [
  {
    "passage_embedding": "$.inference_results.*.output.*.data"
  }
]

To configure an ml_inference search response processor with a local model, specify the function_name explicitly. In this example, the function_name is text_embedding. For information about valid function_name values, see Request fields.

The following is the final configuration of the ml_inference search response processor with the local model:

PUT /_search/pipeline/ml_inference_pipeline_local
{
  "description": "search passage and generates embeddings",
  "processors": [
    {
      "ml_inference": {
        "function_name": "text_embedding",
        "full_response_path": true,
        "model_id": "<your model id>",
        "model_config": {
          "return_number": true,
          "target_response": ["sentence_embedding"]
        },
        "model_input": "{ \"text_docs\": ${input_map.text_docs}, \"return_number\": ${model_config.return_number}, \"target_response\": ${model_config.target_response} }",
        "input_map": [
          {
            "text_docs": "passage_text"
          }
        ],
        "output_map": [
          {
            "passage_embedding": "$.inference_results.*.output.*.data"
          }
        ],
        "ignore_missing": true,
        "ignore_failure": true
      }
    }
  ]
}

Step 2: Run the pipeline

Run the following query, providing the pipeline name in the request:

GET /my_index/_search?search_pipeline=ml_inference_pipeline_local
{
"query": {
  "term": {
    "passage_text": {
      "value": "hello"
      }
    }
  }
}

Response

The response confirms that the processor has generated text embeddings in the passage_embedding field:

{
  "took": 288,
  "timed_out": false,
  "_shards": {
    "total": 1,
    "successful": 1,
    "skipped": 0,
    "failed": 0
  },
  "hits": {
    "total": {
      "value": 1,
      "relation": "eq"
    },
    "max_score": 0.00009405752,
    "hits": [
      {
        "_index": "my_index",
        "_id": "1",
        "_score": 0.00009405752,
        "_source": {
          "passage_text": "hello world",
          "passage_embedding": [
            0.017304314,
            -0.021530833,
            0.050184276,
            0.08962978,
            ...]
        }
      }
    ]
  }
}

Example: Externally hosted text embedding model

The following example shows you how to configure an ml_inference search response processor with an externally hosted model.

Step 1: Create a pipeline

The following example shows you how to create a search pipeline for an externally hosted text embedding model. The model requires an input field and generates results in a data field. It converts the text in the passage_text field into text embeddings and stores the embeddings in the passage_embedding field. The function_name is not explicitly specified in the processor configuration, so it defaults to remote, signifying an externally hosted model:

PUT /_search/pipeline/ml_inference_pipeline
{
  "description": "Generate passage_embedding when search documents",
  "processors": [
    {
      "ml_inference": {
        "model_id": "<your model id>",
        "input_map": [
          {
            "input": "passage_text"
          }
        ],
        "output_map": [
          {
            "passage_embedding": "data"
          }
        ]
      }
    }
  ]
}

When making a Predict API request to an externally hosted model, all necessary fields and parameters are usually contained within a parameters object:

POST /_plugins/_ml/models/cleMb4kBJ1eYAeTMFFg4/_predict
{
  "parameters": {
    "input": [
      {
        ...
      }
    ]
  }
}

When specifying the input_map for an externally hosted model, you can directly reference the input field instead of providing its dot path parameters.input:

"input_map": [
  {
    "input": "passage_text"
  }
]

Step 2: Run the pipeline

Run the following query, providing the pipeline name in the request:

GET /my_index/_search?search_pipeline=ml_inference_pipeline_local
{
  "query": {
    "match_all": {
    }
  }
}

The response confirms that the processor has generated text embeddings in the passage_embedding field. The document within _source now contains both the passage_text and passage_embedding fields:

{
  "took": 288,
  "timed_out": false,
  "_shards": {
    "total": 1,
    "successful": 1,
    "skipped": 0,
    "failed": 0
  },
  "hits": {
    "total": {
      "value": 1,
      "relation": "eq"
    },
    "max_score": 0.00009405752,
    "hits": [
      {
        "_index": "my_index",
        "_id": "1",
        "_score": 0.00009405752,
        "_source": {
          "passage_text": "hello world",
          "passage_embedding": [
            0.017304314,
            -0.021530833,
            0.050184276,
            0.08962978,
            ...]
        }
      }
      }
    ]
  }
}

Example: Externally hosted large language model

This example demonstrates how to configure an ml_inference search response processor to work with an externally hosted large language model (LLM) and map the model’s response to the search extension object. Using the ml_inference processor, you can enable an LLM to summarize search results directly within the response. The summary is included in the ext field of the search response, providing seamless access to AI-generated insights alongside the original search results.

Prerequisite

You must configure an externally hosted LLM for this use case. For more information about externally hosted models, see Connecting to externally hosted models. Once you register the LLM, you can use the following request to test it. This request requires providing the prompt and context fields:

POST /_plugins/_ml/models/KKne6JIBAs32TwoK-FFR/_predict
{
  "parameters": {
    "prompt":"\n\nHuman: You are a professional data analysist. You will always answer question: Which month had the lowest customer acquisition cost per new customer? based on the given context first. If the answer is not directly shown in the context, you will analyze the data and find the answer. If you don't know the answer, just say I don't know. Context: ${parameters.context.toString()}. \n\n Assistant:",
    "context":"Customer acquisition cost: January: $50, February: $45, March: $40. New customers: January: 500, February: 600, March: 750"
  }
}

The response contains the model output in the inference_results field:

{
  "inference_results": [
    {
      "output": [
        {
          "name": "response",
          "dataAsMap": {
            "response": """ Based on the data provided:

                        - Customer acquisition cost in January was $50 and new customers were 500. So cost per new customer was $50/500 = $0.10
                        - Customer acquisition cost in February was $45 and new customers were 600. So cost per new customer was $45/600 = $0.075
                        - Customer acquisition cost in March was $40 and new customers were 750. So cost per new customer was $40/750 = $0.053
            
                        Therefore, the month with the lowest customer acquisition cost per new customer was March, at $0.053."""
          }
        }
      ],
      "status_code": 200
    }
  ]
}

Step 1: Create a pipeline

Create a search pipeline for the registered model. The model requires a context field as input. The model response summarizes the text in the review field and stores the summary in the ext.ml_inference.llm_response field of the search response:

PUT /_search/pipeline/my_pipeline_request_review_llm
{
  "response_processors": [
    {
      "ml_inference": {
        "tag": "ml_inference",
        "description": "This processor is going to run llm",
        "model_id": "EOF6wJIBtDGAJRTD4kNg",
        "function_name": "REMOTE",
        "input_map": [
          {
            "context": "review"
          }
        ],
        "output_map": [
          {
            "ext.ml_inference.llm_response": "response"
          }
        ],
        "model_config": {
          "prompt": "\n\nHuman: You are a professional data analysist. You will always answer question: Which month had the lowest customer acquisition cost per new customer? based on the given context first. If the answer is not directly shown in the context, you will analyze the data and find the answer. If you don't know the answer, just say I don't know. Context: ${parameters.context.toString()}. \n\n Assistant:"
        },
        "ignore_missing": false,
        "ignore_failure": false
      }
    }
  ]
}

In this configuration, you’ve provided the following parameters:

  • The model_id parameter specifies the ID of the generative AI model.
  • The function_name parameter is set to REMOTE, indicating that the model is hosted externally.
  • The input_map parameter maps the review field from the document to the context field expected by the model.
  • The output_map parameter specifies that the model’s response should be stored in ext.ml_inference.llm_response in the search response.
  • The model_config parameter includes a prompt that tells the model how to process the input and generate a summary.

Step 2: Index sample documents

Index some sample documents to test the pipeline:

POST /_bulk
{"index":{"_index":"review_string_index","_id":"1"}}
{"review":"Customer acquisition cost: January: $50, New customers: January: 500."}
{"index":{"_index":"review_string_index","_id":"2"}}
{"review":"Customer acquisition cost: February: $45, New customers: February: 600."}
{"index":{"_index":"review_string_index","_id":"3"}}
{"review":"Customer acquisition cost: March: $40, New customers: March: 750."}

Step 3: Run the pipeline

Run a search query using the pipeline:

GET /review_string_index/_search?search_pipeline=my_pipeline_request_review_llm
{
  "query": {
    "match_all": {}
  }
}

The response includes the original documents and the generated summary in the ext.ml_inference.llm_response field:

{
  "took": 1,
  "timed_out": false,
  "_shards": {
    "total": 1,
    "successful": 1,
    "skipped": 0,
    "failed": 0
  },
  "hits": {
    "total": {
      "value": 3,
      "relation": "eq"
    },
    "max_score": 1,
    "hits": [
      {
        "_index": "review_string_index",
        "_id": "1",
        "_score": 1,
        "_source": {
          "review": "Customer acquisition cost: January: $50, New customers: January: 500."
        }
      },
      {
        "_index": "review_string_index",
        "_id": "2",
        "_score": 1,
        "_source": {
          "review": "Customer acquisition cost: February: $45, New customers: February: 600."
        }
      },
      {
        "_index": "review_string_index",
        "_id": "3",
        "_score": 1,
        "_source": {
          "review": "Customer acquisition cost: March: $40, New customers: March: 750."
        }
      }
    ]
  },
  "ext": {
    "ml_inference": {
      "llm_response": """ Based on the context provided:

      - Customer acquisition cost in January was $50 and new customers were 500. So the cost per new customer was $50/500 = $0.10

      - Customer acquisition cost in February was $45 and new customers were 600. So the cost per new customer was $45/600 = $0.075

      - Customer acquisition cost in March was $40 and new customers were 750. So the cost per new customer was $40/750 = $0.053

      Therefore, the month with the lowest customer acquisition cost per new customer was March, as it had the lowest cost per customer of $0.053."""
    }
  }
}

Example: Reranking search results using a text similarity model

The following example shows you how to configure an ml_inference search response processor with a text similarity model.

Prerequisite

You must configure an externally hosted text similarity model for this use case. For more information about externally hosted models, see Connecting to externally hosted models. Once you register the text similarity model, you can use the following request to test it. This request requires that you provide the text and text_pair fields within the inputs field:

POST /_plugins/_ml/models/Ialx65IBAs32TwoK1lXf/_predict
{
  "parameters": {
    "inputs":
    {
      "text": "I like you",
      "text_pair": "I hate you"
    }
  }
}

The model returns similarity scores for each input document:

{
  "inference_results": [
    {
      "output": [
        {
          "name": "response",
          "dataAsMap": {
            "label": "LABEL_0",
            "score": 0.022704314440488815
          }
        }
      ],
      "status_code": 200
    }
  ]
}

Step 1: Index sample documents

Create an index and add some sample documents:

POST _bulk
{"index":{"_index":"demo-index-0","_id":"1"}}
{"diary":"I hate you"}
{"index":{"_index":"demo-index-0","_id":"2"}}
{"diary":"I love you"}
{"index":{"_index":"demo-index-0","_id":"3"}}
{"diary":"I dislike you"}

Step 2: Create a search pipeline

For this example, you’ll create a search pipeline that uses a text similarity model in a one-to-one inference mode, processing each document in the search results individually. This setup allows the model to make one prediction request per document, providing specific relevance insights for each search hit. When using input_map to map the search request to query text, the JSON path must start with $._request or _request:

PUT /_search/pipeline/my_rerank_pipeline
{
  "response_processors": [
    {
      "ml_inference": {
        "tag": "ml_inference",
        "description": "This processor runs ml inference during search response",
        "model_id": "Ialx65IBAs32TwoK1lXf",
        "model_input":"""{"parameters":{"inputs":{"text":"${input_map.text}","text_pair":"${input_map.text_pair}"}}}""",
        "function_name": "REMOTE",
        "input_map": [
          {
            "text": "diary",
            "text_pair":"$._request.query.term.diary.value"
          }
        ],
        "output_map": [
          {
            "rank_score": "$.score"
          }
        ],
        "full_response_path": false,
        "model_config": {},
        "ignore_missing": false,
        "ignore_failure": false,
        "one_to_one": true
        },
        "rerank": {
          "by_field": {
            "target_field": "rank_score",
            "remove_target_field": true
          }
        }
    }
  ]
}

In this configuration, you’ve provided the following parameters:

  • The model_id parameter specifies the unique identifier of the text similarity model.
  • The function_name parameter is set to REMOTE, indicating that the model is hosted externally.
  • The input_map parameter maps the diary field from each document to the text input of the model as well as the search query term to the text_pair input.
  • The output_map parameter maps the model’s score to a field named rank_score in each document.
  • The model_input parameter formats the input for the model, ensuring that it matches the structure expected by the Predict API.
  • The one_to_one parameter is set to true, ensuring that the model processes each document individually rather than batching multiple documents together.
  • The ignore_missing parameter is set to false, causing the processor to fail if the mapped fields are missing from a document.
  • The ignore_failure parameter is set to false, causing the entire pipeline to fail if the ML inference processor encounters an error.

The rerank processor is applied after ML inference. It reorders the documents based on the rank_score field generated by the ML model and then removes this field from the final results.

Step 3: Run the pipeline

Now perform a search using the created pipeline:

GET /demo-index-0/_search?search_pipeline=my_rerank_pipeline
{
  "query": {
    "term": {
      "dairy": {
        "value": "today"
      }
    }
  }
}

The response includes the original documents and their reranked scores:

{
  "took": 2,
  "timed_out": false,
  "_shards": {
    "total": 1,
    "successful": 1,
    "skipped": 0,
    "failed": 0
  },
  "hits": {
    "total": {
      "value": 3,
      "relation": "eq"
    },
    "max_score": 0.040183373,
    "hits": [
      {
        "_index": "demo-index-0",
        "_id": "1",
        "_score": 0.040183373,
        "_source": {
          "diary": "I hate you"
        }
      },
      {
        "_index": "demo-index-0",
        "_id": "2",
        "_score": 0.022628736,
        "_source": {
          "diary": "I love you"
        }
      },
      {
        "_index": "demo-index-0",
        "_id": "3",
        "_score": 0.0073115323,
        "_source": {
          "diary": "I dislike you"
        }
      }
    ]
  },
  "profile": {
    "shards": []
  }
}

Next steps

350 characters left

Have a question? .

Want to contribute? or .