Automatisation des tests LLM : Évaluation GPT-4 et suivi sur Google Sheets

• •Point de terminaison HTTP Webhook (généré automatiquement par n8n)
• •Peut nécessiter les informations d'identification d'authentification de l'API cible
• •Informations d'identification Google Sheets API

{
  "meta": {
    "instanceId": "45e293393b5dd8437fb351e5b1ef5511ef67e6e0826a1c10b9b68be850b67593"
  },
  "nodes": [
    {
      "id": "2dbc4a8a-4fb6-4679-9d96-2724f79fbac1",
      "name": "Fusion",
      "type": "n8n-nodes-base.merge",
      "position": [
        1980,
        600
      ],
      "parameters": {
        "mode": "combine",
        "options": {},
        "combineBy": "combineByPosition"
      },
      "typeVersion": 3.1
    },
    {
      "id": "146a6af3-58ec-4555-9202-3ce87a83af28",
      "name": "Analyseur de sortie structurée",
      "type": "@n8n/n8n-nodes-langchain.outputParserStructured",
      "position": [
        1540,
        520
      ],
      "parameters": {
        "jsonSchemaExample": "{\n  \"reasoning\": \"The Assistant fabricated a $1 million figure and a 12-month provision that are not found in the source. This breaches factual correctness and completeness. The output would mislead business stakeholders if used without correction.\",\n  \"decision\": \"Fail\"\n}"
      },
      "typeVersion": 1.2
    },
    {
      "id": "83da8236-e5fb-4847-8033-6559f575c7ff",
      "name": "Mettre à jour les résultats",
      "type": "n8n-nodes-base.googleSheets",
      "position": [
        960,
        200
      ],
      "parameters": {
        "columns": {
          "value": {
            "ID": "={{ $json.ID }}",
            "Input": "={{ $json.Input }}",
            "Output": "={{ $json.Output }}",
            "Decision": "={{ $json.output.decision }}",
            "Test No.": "={{ $json[\"Test No\"][\"\"] }}",
            "Reasoning": "={{ $json.output.reasoning }}",
            "AI Platform": "={{ $json[\"AI Platform\"] }}",
            "Reference Answer": "={{ $json[\"Reference Answer\"] }}"
          },
          "schema": [
            {
              "id": "ID",
              "type": "string",
              "display": true,
              "removed": false,
              "required": false,
              "displayName": "ID",
              "defaultMatch": false,
              "canBeUsedToMatch": true
            },
            {
              "id": "Test No.",
              "type": "string",
              "display": true,
              "required": false,
              "displayName": "Test No.",
              "defaultMatch": false,
              "canBeUsedToMatch": true
            },
            {
              "id": "AI Platform",
              "type": "string",
              "display": true,
              "required": false,
              "displayName": "AI Platform",
              "defaultMatch": false,
              "canBeUsedToMatch": true
            },
            {
              "id": "Input",
              "type": "string",
              "display": true,
              "required": false,
              "displayName": "Input",
              "defaultMatch": false,
              "canBeUsedToMatch": true
            },
            {
              "id": "Output",
              "type": "string",
              "display": true,
              "required": false,
              "displayName": "Output",
              "defaultMatch": false,
              "canBeUsedToMatch": true
            },
            {
              "id": "Reference Answer",
              "type": "string",
              "display": true,
              "required": false,
              "displayName": "Reference Answer",
              "defaultMatch": false,
              "canBeUsedToMatch": true
            },
            {
              "id": "Decision",
              "type": "string",
              "display": true,
              "required": false,
              "displayName": "Decision",
              "defaultMatch": false,
              "canBeUsedToMatch": true
            },
            {
              "id": "Reasoning",
              "type": "string",
              "display": true,
              "required": false,
              "displayName": "Reasoning",
              "defaultMatch": false,
              "canBeUsedToMatch": true
            }
          ],
          "mappingMode": "defineBelow",
          "matchingColumns": [
            "ID"
          ],
          "attemptToConvertTypes": false,
          "convertFieldsToString": false
        },
        "options": {},
        "operation": "appendOrUpdate",
        "sheetName": {
          "__rl": true,
          "mode": "list",
          "value": 537199982,
          "cachedResultUrl": "https://docs.google.com/spreadsheets/d/1c73be3fHkKr0DVJYIt9qlNfJcfuUV6DTShp93fa55Ig/edit#gid=537199982",
          "cachedResultName": "Results"
        },
        "documentId": {
          "__rl": true,
          "mode": "url",
          "value": "https://docs.google.com/spreadsheets/d/1c73be3fHkKr0DVJYIt9qlNfJcfuUV6DTShp93fa55Ig/edit?usp=sharing"
        }
      },
      "credentials": {
        "googleSheetsOAuth2Api": {
          "id": "04iXS2lwUVyzn6F2",
          "name": "Google Sheets account"
        }
      },
      "typeVersion": 4.5
    },
    {
      "id": "824c06fb-9104-4c65-a77f-33db0167c0f6",
      "name": "Note adhésive4",
      "type": "n8n-nodes-base.stickyNote",
      "position": [
        560,
        -20
      ],
      "parameters": {
        "color": 4,
        "height": 720,
        "content": "## 2. Execute Subworkflow\nThis node runs immediately (batching requests), but waits for the result before moving to the next step."
      },
      "typeVersion": 1
    },
    {
      "id": "3a20e99f-b183-4362-b909-2fffdd48d0d2",
      "name": "Note adhésive8",
      "type": "n8n-nodes-base.stickyNote",
      "position": [
        -680,
        160
      ],
      "parameters": {
        "width": 460,
        "height": 280,
        "content": "## Data format\nOur Tests Sheet contains the following columns:\n- ID: A unique identifier for each row\n- Test No.: The test that the LLM was given\n- AI Platform: The LLM that was given the test.\n- Input: The input prompt that the LLM was given.\n- Output: The response that the LLM gave.\n- Reference Answer: The \"gold standard\" answer to the input in question, showing how the LLM is expected to respond."
      },
      "typeVersion": 1
    },
    {
      "id": "16fe7cb7-ca24-40f1-855b-e1867bf29b56",
      "name": "Note adhésive9",
      "type": "n8n-nodes-base.stickyNote",
      "position": [
        0,
        -20
      ],
      "parameters": {
        "color": 6,
        "width": 360,
        "height": 180,
        "content": "## 1. Fetch test cases\nWe start by grabbing our list of test cases stored in a Google Sheet [here](https://docs.google.com/spreadsheets/d/1c73be3fHkKr0DVJYIt9qlNfJcfuUV6DTShp93fa55Ig/edit?usp=sharing).\n\nTo start the workflow, you should click \"Execute workflow\" button to the left of the Manual Trigger node."
      },
      "typeVersion": 1
    },
    {
      "id": "86f611e8-ca94-4b9f-a858-45d0fcbdfcfa",
      "name": "Note adhésive15",
      "type": "n8n-nodes-base.stickyNote",
      "position": [
        900,
        -20
      ],
      "parameters": {
        "color": 6,
        "width": 260,
        "height": 180,
        "content": "## 4. Update results\nWe create a new row in our output sheet, containing our original data together with the judge decision/reasoning."
      },
      "typeVersion": 1
    },
    {
      "id": "caa54653-920b-4d4f-abb6-bab54c64350b",
      "name": "Note adhésive16",
      "type": "n8n-nodes-base.stickyNote",
      "position": [
        1320,
        -20
      ],
      "parameters": {
        "color": 4,
        "width": 360,
        "height": 340,
        "content": "## 3. Judge LLM outputs\nOur prompt judges the LLM input/output and decides if the LLM passed the test, based on how well the output fits to the reference answer. \n\nWe also ask for a reason why the judge made its decision, which we can use to refine our eval later.\n\nWe're using OpenRouter here, which lets us easily tweak which LLM we want to use.\n\nThe output parser makes sure that the output is in JSON format, making the data easy to parse in the next step."
      },
      "typeVersion": 1
    },
    {
      "id": "9b22fb78-d6fa-4dad-a543-1b02828d2f2e",
      "name": "Limite",
      "type": "n8n-nodes-base.limit",
      "disabled": true,
      "position": [
        360,
        220
      ],
      "parameters": {
        "maxItems": 3
      },
      "typeVersion": 1
    },
    {
      "id": "faad2c18-defc-4644-b9a3-3650c26f5891",
      "name": "Extraire les données",
      "type": "n8n-nodes-base.set",
      "position": [
        1000,
        400
      ],
      "parameters": {
        "mode": "raw",
        "options": {},
        "jsonOutput": "={{ $json.body }}"
      },
      "typeVersion": 3.4
    },
    {
      "id": "ec8629e4-7715-410c-aa6d-560fd284a1ca",
      "name": "Obtenir les tests",
      "type": "n8n-nodes-base.googleSheets",
      "position": [
        140,
        220
      ],
      "parameters": {
        "options": {},
        "sheetName": {
          "__rl": true,
          "mode": "list",
          "value": "gid=0",
          "cachedResultUrl": "https://docs.google.com/spreadsheets/d/1c73be3fHkKr0DVJYIt9qlNfJcfuUV6DTShp93fa55Ig/edit#gid=0",
          "cachedResultName": "Tests"
        },
        "documentId": {
          "__rl": true,
          "mode": "url",
          "value": "https://docs.google.com/spreadsheets/d/1c73be3fHkKr0DVJYIt9qlNfJcfuUV6DTShp93fa55Ig/edit?usp=sharing"
        }
      },
      "credentials": {
        "googleSheetsOAuth2Api": {
          "id": "04iXS2lwUVyzn6F2",
          "name": "Google Sheets account"
        }
      },
      "typeVersion": 4.5
    },
    {
      "id": "d7160cac-8bea-4464-bfea-00c785b8ac7e",
      "name": "Exécuter le sous-workflow",
      "type": "n8n-nodes-base.httpRequest",
      "onError": "continueErrorOutput",
      "maxTries": 2,
      "position": [
        620,
        220
      ],
      "parameters": {
        "url": "https://webhook-processor-production-48f8.up.railway.app/webhook/llm-as-a-judge",
        "method": "POST",
        "options": {
          "batching": {
            "batch": {
              "batchSize": 1,
              "batchInterval": 500
            }
          }
        },
        "jsonBody": "={{ $json }}",
        "sendBody": true,
        "specifyBody": "json"
      },
      "retryOnFail": false,
      "typeVersion": 4.2
    },
    {
      "id": "6920a43b-bdbf-47c0-a644-1f75375e1127",
      "name": "Webhook",
      "type": "n8n-nodes-base.webhook",
      "position": [
        620,
        480
      ],
      "webhookId": "1cbce320-d28e-4e97-8663-bf2c6a36a358",
      "parameters": {
        "path": "llm-as-a-judge",
        "options": {},
        "httpMethod": "POST",
        "responseData": "allEntries",
        "responseMode": "lastNode"
      },
      "typeVersion": 2
    },
    {
      "id": "70cc9edd-f481-420e-bfcc-02b25f4353db",
      "name": "Chaîne LLM de base",
      "type": "@n8n/n8n-nodes-langchain.chainLlm",
      "onError": "continueErrorOutput",
      "position": [
        1380,
        340
      ],
      "parameters": {
        "text": "=INPUT:\n\n{\n  \"task\": {{ $('Extract Data').item.json['Input'] }},\n  \"answer_key\": {{ $('Extract Data').item.json['Reference Answer'] }},\n  \"output\": {{ $('Extract Data').item.json['Output'] }}\n}\n\nOUTPUT:",
        "messages": {
          "messageValues": [
            {
              "message": "=## Context\n\nYou are an evaluator of LLMs in the legal domain.\n\n## Inputs Provided for Each Task\n\n- task: The legal question or instruction.\n- answer_key: The correct answer for this task, found in the answer key column of the same Google Sheet.\n- output: The answer generated by the AI Assistant.\n\n\n## Evaluation Rules\n\nGrade the AI Assistant's output as Pass or Fail by comparing it ONLY to the answer_key for that task.\n\nDo not use or reference the original source material or any other information.\n\n## Criteria for Pass\n\n1. Factual Correctness\n- The output must accurately reflect the information in the answer_key.\n- Minor differences in paraphrasing, wording, or formatting (including clause numbering, references, or synonyms) are acceptable if the substantive information matches the answer_key.\n- If the answer key provides multiple possible correct answers (e.g., separated by \"OR\"), any output that matches any one of the alternatives is acceptable.\n\n\n2. Relevance to the Query\n- The output must directly answer the task as covered in the answer_key.\n- Do not introduce unrelated or off-topic information.\n\n\n3. Completeness\n- If the output contains extra information that does not contradict or misrepresent the answer key, it is acceptable.\n- Omitting any critical point present in the answer_key = Fail.\n\n\n## Key Rule\n- If the output materially fails any one of the three requirements compared to the answer_key, grade as Fail.\n- Minor paraphrasing or stylistic differences are acceptable if the substantive meaning is identical.\n\n\n## Required Output Format\n\nYour evaluation must be provided in JSON with two keys only:\n\n- decision: Pass or Fail\n- reasoning: A brief explanation, strictly comparing the output to the answer_key.\n\n\n### Example Input 1\n\n{\n \"task\": \"Extract the liability cap and time-based provisions from a limitation of liability clause.\",\n \"answer_key\": \"The liability cap is $1 million with a 12-month limit.\",\n \"output\": \"The liability cap is $1 million with a 12-month limit.\"\n}\n\n### Example Output 1\n\n{\n  \"output\": {\n    {\n     \"decision\": \"Pass\",\n     \"reasoning\": \"The output exactly matches the answer key, so it is factually correct, relevant, and complete.\"\n     }\n  }\n}\n\n### Example Input 2\n\n{\n \"task\": \"Extract the liability cap and time-based provisions from a limitation of liability clause.\",\n \"answer_key\": \"The liability cap is $1 million with a 12-month limit.\",\n \"output\": \"The liability cap is $2 million and there is no time limit.\"\n}\n\n### Example Output 2\n\n{\n  \"output\": {\n    {\n     \"decision\": \"Fail\",\n     \"reasoning\": \"The output gives a $2 million cap and omits the 12-month limit from the answer key. This fails both factual correctness and completeness.\"\n    }\n}\n\n### Example Input 3\n\n{\n \"task\": \"State the governing law.\",\n \"answer_key\": \"Singapore law.\",\n \"output\": \"This agreement is governed by Singapore law. All disputes will be subject to the exclusive jurisdiction of Singapore courts.\"\n}\n\n### Example Output 3\n\n{\n  \"output\": {\n    \"reasoning\": \"All required information from the answer_key is present. The extra information does not contradict or misrepresent the answer_key.\"\n \"decision\": \"Pass\",\n  }\n}\n\n### Example Input 4\n\n{\n \"task\": \"Identify the relevant clause.\",\n \"answer_key\": \"Clause 5\",\n \"output\": \"clause 5\"\n}\n\n### Example Output 4\n\n{\n  \"output\": {\n    \"reasoning\": \"The output matches the answer key despite minor formatting differences.\"\n    \"decision\": \"Pass\",\n  }\n}\n\n### Example Input 5\n\n{\n \"task\": \"Extract the parties to the contract.\",\n \"answer_key\": \"Company A and Company B OR The Buyer and the Seller\",\n \"output\": \"The Buyer and the Seller\"\n}\n\n### Example Output 5\n\n{\n  \"output\": {\n    \"reasoning\": \"The output matches one of the acceptable answer_key alternatives.\"\n    \"decision\": \"Pass\",\n  }\n}\n\n## Reminder\nAlways grade solely by comparison to the answer_key column for each task in the input data."
            }
          ]
        },
        "promptType": "define",
        "hasOutputParser": true
      },
      "typeVersion": 1.4
    },
    {
      "id": "f4ddb551-cbaa-4c2d-96ca-3769a199ce1a",
      "name": "Modèle de chat OpenRouter",
      "type": "@n8n/n8n-nodes-langchain.lmChatOpenRouter",
      "position": [
        1380,
        520
      ],
      "parameters": {
        "model": "openai/gpt-4.1",
        "options": {}
      },
      "credentials": {
        "openRouterApi": {
          "id": "ipzDVYsZqbum9bX4",
          "name": "OpenRouter account 2"
        }
      },
      "typeVersion": 1
    },
    {
      "id": "b8eedf4a-eb85-4b4a-ad4b-61d9d31984c1",
      "name": "Conserver les données d'origine",
      "type": "n8n-nodes-base.set",
      "position": [
        1480,
        820
      ],
      "parameters": {
        "mode": "raw",
        "options": {},
        "jsonOutput": "={{ $json.body }}"
      },
      "typeVersion": 3.4
    },
    {
      "id": "69c41be1-ff93-4098-8b9d-cd5cc88d9271",
      "name": "Déclencheur manuel",
      "type": "n8n-nodes-base.manualTrigger",
      "position": [
        -80,
        220
      ],
      "parameters": {},
      "typeVersion": 1
    }
  ],
  "pinData": {},
  "connections": {
    "9b22fb78-d6fa-4dad-a543-1b02828d2f2e": {
      "main": [
        [
          {
            "node": "d7160cac-8bea-4464-bfea-00c785b8ac7e",
            "type": "main",
            "index": 0
          }
        ]
      ]
    },
    "6920a43b-bdbf-47c0-a644-1f75375e1127": {
      "main": [
        [
          {
            "node": "b8eedf4a-eb85-4b4a-ad4b-61d9d31984c1",
            "type": "main",
            "index": 0
          },
          {
            "node": "faad2c18-defc-4644-b9a3-3650c26f5891",
            "type": "main",
            "index": 0
          }
        ]
      ]
    },
    "ec8629e4-7715-410c-aa6d-560fd284a1ca": {
      "main": [
        [
          {
            "node": "9b22fb78-d6fa-4dad-a543-1b02828d2f2e",
            "type": "main",
            "index": 0
          }
        ]
      ]
    },
    "faad2c18-defc-4644-b9a3-3650c26f5891": {
      "main": [
        [
          {
            "node": "70cc9edd-f481-420e-bfcc-02b25f4353db",
            "type": "main",
            "index": 0
          }
        ]
      ]
    },
    "69c41be1-ff93-4098-8b9d-cd5cc88d9271": {
      "main": [
        [
          {
            "node": "ec8629e4-7715-410c-aa6d-560fd284a1ca",
            "type": "main",
            "index": 0
          }
        ]
      ]
    },
    "70cc9edd-f481-420e-bfcc-02b25f4353db": {
      "main": [
        [
          {
            "node": "2dbc4a8a-4fb6-4679-9d96-2724f79fbac1",
            "type": "main",
            "index": 0
          }
        ]
      ]
    },
    "b8eedf4a-eb85-4b4a-ad4b-61d9d31984c1": {
      "main": [
        [
          {
            "node": "2dbc4a8a-4fb6-4679-9d96-2724f79fbac1",
            "type": "main",
            "index": 1
          }
        ]
      ]
    },
    "d7160cac-8bea-4464-bfea-00c785b8ac7e": {
      "main": [
        [
          {
            "node": "83da8236-e5fb-4847-8033-6559f575c7ff",
            "type": "main",
            "index": 0
          }
        ]
      ]
    },
    "f4ddb551-cbaa-4c2d-96ca-3769a199ce1a": {
      "ai_languageModel": [
        [
          {
            "node": "70cc9edd-f481-420e-bfcc-02b25f4353db",
            "type": "ai_languageModel",
            "index": 0
          }
        ]
      ]
    },
    "146a6af3-58ec-4555-9202-3ce87a83af28": {
      "ai_outputParser": [
        [
          {
            "node": "70cc9edd-f481-420e-bfcc-02b25f4353db",
            "type": "ai_outputParser",
            "index": 0
          }
        ]
      ]
    }
  }
}