RAG Reasoning

Available in Classic and VPC

Generate evidence-based RAG answers by leveraging RAG Reasoning model trained on answer types, such as citation sources that increase credibility and citation source indexing notations. RAG Reasoning calls the engine in the function calling format. You can specify a single or multiple RAG functions, and LLM will autonomously select the best function for the context to generate search augmentation. When used in conjunction with Reranker, you can get more stable results.

Request

This section describes the request format. The method and URI are as follows:

Request headers

The following describes the request headers.

Request body

You can include the following data in the body of your request:

messages

The following describes messages.

{
    "role": "tool",
    "content": "[
                    {
                        \"search_result\": [{\"id\": \"doc-1493058999\",
                        \"doc\": \"Login with NAVER ID is only available for individual members. Business members can't use this feature.\"
                    },
                    ...
                 ]"
}

tools

The following describes tools.

toolCalls

The following describes toolCalls.

Request example

The request example is as follows:

• Step 1. Enter your query in role: user and call the best function to generate the answer (Check response)

  curl --location --request POST 'https://clovastudio.stream.ntruss.com/v1/api-tools/rag-reasoning' \
  --header 'Authorization: Bearer <access_token>' \
  --header 'Content-Type: application/json' \
  --data-raw '{
    "messages": [
      {
        "content": "How to rent an A100 GPU",
        "role": "user"
      }
    ],
    "tools": [
      {
        "function": {
          "description": "This is the tool you use to do Ncloud-related searches.\nUse the tool by breaking up your query if you need to ask multiple questions.\nIf you can't find information, you can use the tool again with suggested_queries as a reference without giving a final answer.",
          "name": "ncloud_cs_retrieval",
          "parameters": {
            "properties": {
              "query": {
                "description": "Refine and enter the user's search keywords.",
                "type": "string"
              }
            },
            "required": [
              "query"
            ],
            "type": "object"
          }
        },
        "type": "function"
      }
    ],
    "toolChoice": "auto",
    "maxTokens": 1024
  }'
  

• Step 2. Request with role: tool to generate the final answer (Check response)

  curl --location --request POST 'https://clovastudio.stream.ntruss.com/v1/api-tools/rag-reasoning' \
  --header 'Authorization: Bearer <access_token>' \
  --header 'Content-Type: application/json' \
  --data-raw '{
      "messages": [
          {
              "content": "How to rent an A100 GPU",
              "role": "user"
          },
          {
              "role": "assistant",
              "content": "",
              "toolCalls": [
                  {
                      "id": "call_enTEYb0kWBjOwtkngbl7FGTm",
                      "type": "function",
                      "function": {
                          "name": "ncloud_cs_retrieval",
                          "arguments": {
                              "query": "How to rent an A100 GPU"
                          }
                      }
                  }
              ]
          },
          {
              "content": "{\"search_result\": [{\"id\": \"doc-179\", \"doc\": \"GPU A100 can only be created in KR-1. When creating an A100, select a subnet in KR-1. Up to 5 GPU servers can be created for corporate members only.\"}, {\"id\": \"doc-248\", \"doc\": \"GPU A100 servers can be created in the Services > Compute > Server menu. For more information, see the Create server guide.\"}, {\"id\": \"doc-156\", \"doc\": \"For individual members who need more GPU servers or need to create a GPU server, please refer to the FAQ and contact Support.\"}]}",
              "role": "tool",
              "toolCallId": "call_enTEYb0kWBjOwtkngbl7FGTm"
          }
      ],
      "tools": [
          {
              "function": {
                  "description": "This is the tool you use to do Ncloud-related searches.\nUse the tool by breaking up your query if you need to ask multiple questions.\nIf you can't find information, you can use the tool again with suggested_queries as a reference without giving a final answer.",
                  "name": "ncloud_cs_retrieval",
                  "parameters": {
                      "properties": {
                          "query": {
                              "description": "Refine and enter the user's search keywords.",
                              "type": "string"
                          }
                      },
                      "required": [
                          "query"
                      ],
                      "type": "object"
                  }
              },
              "type": "function"
          }
      ]
  }'
  

Response

This section describes the response format.

Response headers

The following describes the response headers.

Response body

The response body includes the following data:

toolCalls

The following describes toolCalls.

Response example

The response example is as follows:

• Example response to Step 1. (Check response)

  {
      "status": {
          "code": "20000",
          "message": "OK"
      },
      "result": {
          "message": {
              "role": "assistant",
              "content": "",
              "thinkingContent": "The user has inquired about \"how to rent an A100 GPU\". To find the answer to this question, you need to use the tool "ncloud_cs_retrieval" to retrieve relevant information.",
              "toolCalls": [
                  {
                      "id": "call_enTEYb0kWBjOwtkngbl7FGTm",
                      "type": "function",
                      "function": {
                          "name": "ncloud_cs_retrieval",
                          "arguments": {
                              "query": "How to rent an A100 GPU"
                          }
                      }
                  }
              ]
          },
          "usage": {
              "promptTokens": 135,
              "completionTokens": 84,
              "totalTokens": 219
          }
      }
  }
  

• Example response to Step 2. (LLM returns the final answer) (Check response)

  {
      "status": {
          "code": "20000",
          "message": "OK"
      },
      "result": {
          "message": {
              "role": "assistant",
              "content": "To rent an A100 GPU, <doc-248>you can create a GPU A100 server from the Services > Compute > Server menu in the NAVER Cloud Platform console.</doc-248>However, <doc-179>GPU A100 can only be created in KR-1, and you must select a subnet in KR-1 when creating an A100.</doc-179> Also, <doc-179>up to GPU servers can be created for corporate members only.</doc-179> If you need more GPU servers or are an individual member who needs to create a GPU server, <doc-156>please refer to the FAQ and contact Support.</doc-156>"
          },
          "usage": {
              "promptTokens": 332,
              "completionTokens": 146,
              "totalTokens": 478
          }
      }
  }
  

Failure

The following is a sample response upon a failed call.

• Common client error troubleshooting (4xx)
• Common server error troubleshooting (5xx)