Qdrant と n8n を使用したハイブリッド検索、法務 AI：検索
上級
これは自動化ワークフローで、17個のノードを含みます。主にSet, Merge, Filter, SplitOut, Qdrantなどのノードを使用。 Qdrantとn8nに基づいたハイブリッド検索、法務AI：検索
前提条件
- •Qdrantサーバー接続情報
- •ターゲットAPIの認証情報が必要な場合あり
カテゴリー
-
ワークフロープレビュー
ノード接続関係を可視化、ズームとパンをサポート
ワークフローをエクスポート
以下のJSON設定をn8nにインポートして、このワークフローを使用できます
{
"id": "h81ddl7uooV3eLBq",
"meta": {
"instanceId": "d975180a7308eb9e1d0eb6c8833136580b02ced551ba46ad477d3b76dff98527",
"templateCredsSetupCompleted": true
},
"name": "Hybrid Search with Qdrant & n8n, Legal AI: Retrieval",
"tags": [],
"nodes": [
{
"id": "eb8d4dd7-f40b-4524-a9de-f9ef9eef0eca",
"name": "HuggingFaceからのデータセットインデックス作成",
"type": "n8n-nodes-base.manualTrigger",
"position": [
-256,
400
],
"parameters": {},
"typeVersion": 1
},
{
"id": "78030f66-5331-463f-ad22-9d09f477e3f9",
"name": "一括分割",
"type": "n8n-nodes-base.splitOut",
"position": [
176,
400
],
"parameters": {
"options": {},
"fieldToSplitOut": "splits"
},
"typeVersion": 1
},
{
"id": "e6d1e789-1293-480b-a163-992b0c7a2ae8",
"name": "データセット分割の取得",
"type": "n8n-nodes-base.httpRequest",
"position": [
-32,
400
],
"parameters": {
"url": "https://datasets-server.huggingface.co/splits",
"options": {},
"sendQuery": true,
"queryParameters": {
"parameters": [
{
"name": "dataset",
"value": "={{ $json.dataset }}"
}
]
}
},
"typeVersion": 4.2
},
{
"id": "cb58241e-4579-4c5b-bd65-4a20f6cf3698",
"name": "行ごとに分割",
"type": "n8n-nodes-base.splitOut",
"position": [
816,
400
],
"parameters": {
"options": {},
"fieldToSplitOut": "rows"
},
"typeVersion": 1
},
{
"id": "2ee71c28-71fd-4cba-87c7-c9886fb403c7",
"name": "テスト分割を保持",
"type": "n8n-nodes-base.filter",
"position": [
384,
400
],
"parameters": {
"options": {},
"conditions": {
"options": {
"version": 2,
"leftValue": "",
"caseSensitive": true,
"typeValidation": "strict"
},
"combinator": "and",
"conditions": [
{
"id": "52e3d8e2-825f-4e43-9d5f-e275d196b442",
"operator": {
"name": "filter.operator.equals",
"type": "string",
"operation": "equals"
},
"leftValue": "={{ $json.split }}",
"rightValue": "test"
}
]
}
},
"typeVersion": 2.2
},
{
"id": "484a82fe-93d8-439b-bbb5-e96a4b5d7861",
"name": "テストクエリを取得",
"type": "n8n-nodes-base.httpRequest",
"position": [
592,
400
],
"parameters": {
"url": "=https://datasets-server.huggingface.co/rows",
"options": {},
"sendQuery": true,
"queryParameters": {
"parameters": [
{
"name": "dataset",
"value": "={{ $json.dataset }}"
},
{
"name": "config",
"value": "={{ $json.config }}"
},
{
"name": "split",
"value": "={{ $json.split }}"
},
{
"name": "length",
"value": "=100"
}
]
}
},
"typeVersion": 4.2
},
{
"id": "20f67ae7-6631-4602-aa20-42a382db12ae",
"name": "ポイントをクエリ",
"type": "n8n-nodes-qdrant.qdrant",
"position": [
2144,
416
],
"parameters": {
"limit": 1,
"query": "{\n \"fusion\": \"rrf\"\n}",
"prefetch": "=[\n {\n \"query\": {\n \"text\": \"{{ $json.question }}\",\n \"model\": \"mixedbread-ai/mxbai-embed-large-v1\"\n },\n \"using\": \"mxbai_large\",\n \"limit\": 25\n },\n {\n \"query\": {\n \"text\": \"{{ $json.question }}\",\n \"model\": \"qdrant/bm25\"\n },\n \"using\": \"bm25\",\n \"limit\": 25\n }\n]",
"resource": "search",
"operation": "queryPoints",
"collectionName": {
"__rl": true,
"mode": "list",
"value": "legalQA_test",
"cachedResultName": "legalQA_test"
},
"requestOptions": {}
},
"credentials": {
"qdrantApi": {
"id": "LVjhdCt8pAJjLyt5",
"name": "Qdrant account 2"
}
},
"typeVersion": 1
},
{
"id": "445ace25-f900-4bcf-9f7d-9bc1db662867",
"name": "マージ",
"type": "n8n-nodes-base.merge",
"position": [
2320,
608
],
"parameters": {
"mode": "combine",
"options": {},
"combineBy": "combineAll"
},
"typeVersion": 3.2
},
{
"id": "c631ce99-a672-499f-bbf3-e740ef431884",
"name": "アイテムをループ処理",
"type": "n8n-nodes-base.splitInBatches",
"position": [
1776,
400
],
"parameters": {
"options": {
"reset": false
}
},
"typeVersion": 3
},
{
"id": "c075a745-ed50-459f-87dd-101a559e4523",
"name": "データセット内に回答がある質問を保持",
"type": "n8n-nodes-base.filter",
"position": [
1056,
400
],
"parameters": {
"options": {},
"conditions": {
"options": {
"version": 2,
"leftValue": "",
"caseSensitive": true,
"typeValidation": "strict"
},
"combinator": "and",
"conditions": [
{
"id": "d1120153-1852-42c0-8b0a-084e8c3190d3",
"operator": {
"type": "number",
"operation": "gt"
},
"leftValue": "={{ $json.row.answers.length }}",
"rightValue": 0
}
]
}
},
"typeVersion": 2.2
},
{
"id": "94f78e9b-f9eb-4179-9844-d6e23bc79751",
"name": "質問とIDを保持",
"type": "n8n-nodes-base.set",
"position": [
1280,
400
],
"parameters": {
"options": {},
"assignments": {
"assignments": [
{
"id": "961c95d9-c803-404b-b4b6-cb66a8a33928",
"name": "id_qa",
"type": "string",
"value": "={{ $json.row.id }}"
},
{
"id": "0fefba06-4567-479c-9eb5-efbb3e13e743",
"name": "question",
"type": "string",
"value": "={{ $json.row.question }}"
}
]
}
},
"typeVersion": 3.4
},
{
"id": "08f67e31-ba8f-47fd-bc78-f352a160d4fd",
"name": "評価を集計",
"type": "n8n-nodes-base.aggregate",
"position": [
2032,
224
],
"parameters": {
"options": {},
"aggregate": "aggregateAllItemData",
"destinationFieldName": "eval"
},
"typeVersion": 1
},
{
"id": "413ade44-d27d-4c49-862f-afb9d4e18bf6",
"name": "評価内のisHitsの割合",
"type": "n8n-nodes-base.set",
"position": [
2256,
224
],
"parameters": {
"options": {},
"assignments": {
"assignments": [
{
"id": "5bca1a50-3e41-4f50-8362-cb7b185b50f6",
"name": "Hits percentage",
"type": "number",
"value": "={{ ($json.eval.filter(item => item.isHit).length * 100) / $json.eval.length}}"
}
]
}
},
"typeVersion": 3.4
},
{
"id": "c0840c22-8954-4937-80d9-f32741b81e1e",
"name": "付箋2",
"type": "n8n-nodes-base.stickyNote",
"position": [
-96,
144
],
"parameters": {
"color": 5,
"width": 1520,
"height": 464,
"content": "## Get Questions to Eval Retrieval from Hugging Face Dataset (Already Indexed to Qdrant)\n\nFetching questions from a sample Q&A dataset on Hugging Face using the [Dataset Viewer API](https://huggingface.co/docs/dataset-viewer/quick_start). \n**Dataset:** [LegalQAEval (isaacus)](https://huggingface.co/datasets/isaacus/LegalQAEval)\n\n1. **Retrieve dataset splits**. \n2. **Get a small subsample of questions from the `test` split**. \n To fetch the full split, apply [pagination in HTTP node](https://docs.n8n.io/code/cookbook/http-node/pagination/#enable-pagination), as shown in Part 1. \n3. **Keep only questions that have a paired text chunk answering them**, so evaluation remains fair. \n"
},
"typeVersion": 1
},
{
"id": "742e68ae-0013-4dda-a818-4485ff80a986",
"name": "付箋4",
"type": "n8n-nodes-base.stickyNote",
"position": [
1696,
-256
],
"parameters": {
"color": 5,
"width": 1088,
"height": 1120,
"content": "## Check Quality of Simple Hybrid Search on Legal Q&A Dataset\nFor each question in the evaluation set, using the qdrant collection created and indexed in Part 1:\n1. **Perform a Hybrid Search in Qdrant** \n - Get 25 results with [**BM25-based keyword retrieval**](https://en.wikipedia.org/wiki/Okapi_BM25) (exact word matches). \n - Sparse representations for BM25 are created automatically by Qdrant. \n - Get 25 results with [**mxbai-embed-large-v1**](https://huggingface.co/mixedbread-ai/mxbai-embed-large-v1) semantic search (meaning-based matches). \n - Here we use [**Qdrant Cloud Inference**](https://qdrant.tech/documentation/cloud/inference/), so conversion of questions to vectors and searching is handled by the Qdrant node. \n - To use an external provider (e.g. OpenAI), see Part 1 for an example on how to adapt this template. \n - Fuse both result lists with **Reciprocal Rank Fusion (RRF)**. \n - Select the **top-1 result**. \n2. **Check the top-1 result** \n - Verify if the text chunk contains the correct answer. This is done by checking if the question ID is present in the list of related to the text chunk question IDs (created in Part 1). \n3. **Aggregate results** \n - Calculate the **hits@1**: percentage of evaluation questions where the top-1 retrieved chunk contained the answer. \n\n- If results are good → you can reuse the **Qdrant Query Points** node as a tool for an **agentic legal AI RAG** system. \n- If results are poor → don’t worry. This is the *simplest* hybrid query setup. You can improve quality with [various tooling for hybrid search in Qdrant](https://qdrant.tech/documentation/concepts/hybrid-queries/): \n - Reranking \n - Score boosting \n - Tuning vector index parameters \n - … \n\n\nExperiment! 🙂\n\n"
},
"typeVersion": 1
},
{
"id": "d4a32298-02ef-4f9e-b22d-3f30e9b74eb2",
"name": "付箋1",
"type": "n8n-nodes-base.stickyNote",
"position": [
-1344,
-128
],
"parameters": {
"width": 1008,
"height": 960,
"content": "## Evaluate Hybrid Search on Legal Dataset\n*This is the second part of **\"Hybrid Search with Qdrant & n8n, Legal AI.\"**\nThe first part, **\"Indexing,\"** covers preparing and uploading the dataset to Qdrant.*\n\n### Overview\nThis pipeline demonstrates how to perform **Hybrid Search** on a [Qdrant collection](https://qdrant.tech/documentation/concepts/collections/#collections) using `question`s and `text` chunks (containing answers) from the \n[LegalQAEval dataset (isaacus)](https://huggingface.co/datasets/isaacus/LegalQAEval).\n\nOn a small subset of questions, it shows: \n- How to set up hybrid retrieval in Qdrant with: \n - [BM25](https://en.wikipedia.org/wiki/Okapi_BM25)-based keyword retrieval;\n - [mxbai-embed-large-v1](https://huggingface.co/mixedbread-ai/mxbai-embed-large-v1) semantic retrieval; \n - **Reciprocal Rank Fusion (RRF)**, a simple zero-shot fusion of the two searches;\n- How to run a basic evaluation: \n - Calculate **hits@1** — the percentage of evaluation questions where the top-1 retrieved text chunk contains the correct answer \n\n\nAfter running this pipeline, you will have a quality estimate of a simple hybrid retrieval setup. \nFrom there, you can reuse Qdrant’s **Query Points** node to build a **legal RAG chatbot**. \n\n### Embedding Inference\n- By default, this pipeline uses [**Qdrant Cloud Inference**](https://qdrant.tech/documentation/cloud/inference/) to convert questions to embeddings. \n- You can also use an **external embedding provider** (e.g. OpenAI). \n - In that case, minimally update the pipeline, similar to the adjustments showed in **Part 1: Indexing**. \n\n### Prerequisites\n- **Completed Part 1 pipeline**, *\"Hybrid Search with Qdrant & n8n, Legal AI: Indexing\"*, and the collection created in it;\n- All the requirements of **Part 1 pipeline**;\n\n### Hybrid Search\nThe example here is a **basic hybrid query**. You can extend/enhance it with:\n- Reranking strategies; \n- Different fusion techniques;\n- Score boosting based on metadata;\n- ... \n\nMore details: [Hybrid Queries in Qdrant](https://qdrant.tech/documentation/concepts/hybrid-queries/). \n\n#### P.S.\n- To ask retrieval in Qdrant-related questions, join the [Qdrant Discord](https://discord.gg/ArVgNHV6). \n- Star [Qdrant n8n community node repo](https://github.com/qdrant/n8n-nodes-qdrant) <3\n"
},
"typeVersion": 1
},
{
"id": "56a5efd8-ed3f-46f7-85c8-966536f24a13",
"name": "isHit = 正解が見つかった場合",
"type": "n8n-nodes-base.set",
"position": [
2512,
608
],
"parameters": {
"include": "selected",
"options": {},
"assignments": {
"assignments": [
{
"id": "80089820-cc55-4b74-966e-b50a3f4b6e36",
"name": "isHit",
"type": "boolean",
"value": "={{ $json.result.points[0].payload.ids_qa.includes($json.id_qa) }}"
}
]
},
"includeFields": "id_qa,question",
"includeOtherFields": true
},
"typeVersion": 3.4
}
],
"active": false,
"pinData": {
"Index Dataset from HuggingFace": [
{
"json": {
"dataset": "isaacus/LegalQAEval"
}
}
]
},
"settings": {
"executionOrder": "v1"
},
"versionId": "20b38566-7985-4139-98a3-6b275e85a9cb",
"connections": {
"445ace25-f900-4bcf-9f7d-9bc1db662867": {
"main": [
[
{
"node": "56a5efd8-ed3f-46f7-85c8-966536f24a13",
"type": "main",
"index": 0
}
]
]
},
"20f67ae7-6631-4602-aa20-42a382db12ae": {
"main": [
[
{
"node": "445ace25-f900-4bcf-9f7d-9bc1db662867",
"type": "main",
"index": 0
}
]
]
},
"cb58241e-4579-4c5b-bd65-4a20f6cf3698": {
"main": [
[
{
"node": "c075a745-ed50-459f-87dd-101a559e4523",
"type": "main",
"index": 0
}
]
]
},
"08f67e31-ba8f-47fd-bc78-f352a160d4fd": {
"main": [
[
{
"node": "413ade44-d27d-4c49-862f-afb9d4e18bf6",
"type": "main",
"index": 0
}
]
]
},
"2ee71c28-71fd-4cba-87c7-c9886fb403c7": {
"main": [
[
{
"node": "484a82fe-93d8-439b-bbb5-e96a4b5d7861",
"type": "main",
"index": 0
}
]
]
},
"c631ce99-a672-499f-bbf3-e740ef431884": {
"main": [
[
{
"node": "08f67e31-ba8f-47fd-bc78-f352a160d4fd",
"type": "main",
"index": 0
}
],
[
{
"node": "445ace25-f900-4bcf-9f7d-9bc1db662867",
"type": "main",
"index": 1
},
{
"node": "20f67ae7-6631-4602-aa20-42a382db12ae",
"type": "main",
"index": 0
}
]
]
},
"484a82fe-93d8-439b-bbb5-e96a4b5d7861": {
"main": [
[
{
"node": "cb58241e-4579-4c5b-bd65-4a20f6cf3698",
"type": "main",
"index": 0
}
]
]
},
"e6d1e789-1293-480b-a163-992b0c7a2ae8": {
"main": [
[
{
"node": "78030f66-5331-463f-ad22-9d09f477e3f9",
"type": "main",
"index": 0
}
]
]
},
"78030f66-5331-463f-ad22-9d09f477e3f9": {
"main": [
[
{
"node": "2ee71c28-71fd-4cba-87c7-c9886fb403c7",
"type": "main",
"index": 0
}
]
]
},
"94f78e9b-f9eb-4179-9844-d6e23bc79751": {
"main": [
[
{
"node": "c631ce99-a672-499f-bbf3-e740ef431884",
"type": "main",
"index": 0
}
]
]
},
"eb8d4dd7-f40b-4524-a9de-f9ef9eef0eca": {
"main": [
[
{
"node": "e6d1e789-1293-480b-a163-992b0c7a2ae8",
"type": "main",
"index": 0
}
]
]
},
"56a5efd8-ed3f-46f7-85c8-966536f24a13": {
"main": [
[
{
"node": "c631ce99-a672-499f-bbf3-e740ef431884",
"type": "main",
"index": 0
}
]
]
},
"c075a745-ed50-459f-87dd-101a559e4523": {
"main": [
[
{
"node": "94f78e9b-f9eb-4179-9844-d6e23bc79751",
"type": "main",
"index": 0
}
]
]
}
}
}よくある質問
このワークフローの使い方は？
上記のJSON設定コードをコピーし、n8nインスタンスで新しいワークフローを作成して「JSONからインポート」を選択、設定を貼り付けて認証情報を必要に応じて変更してください。
このワークフローはどんな場面に適していますか？
上級
有料ですか？
このワークフローは完全無料です。ただし、ワークフローで使用するサードパーティサービス（OpenAI APIなど）は別途料金が発生する場合があります。
関連ワークフロー
Qdrant と n8n を使用したハイブリッド検索、法務 AI：インデックス
Qdrantとn8nに基づいたハイブリッド検索、法務AI：インデックス
If
Set
Limit
+
If
Set
Limit
37 ノードJenny
YNABによる自動のな予算管理
GPT-5-Miniを使ってYNABの取引を自動分類しDiscordへの通知を送信
If
Set
Merge
+
If
Set
Merge
29 ノードspencer owen
AI要約
GPT-5 Mini、Jira、そしてフォームインターフェースを使った敏捷開発プロジェクト設定の自動化
GPT-5 Mini、Jira、そしてフォームインターフェースを使った敏捷なプロジェクトセットアップの自動化
Set
Jira
Gmail
+
Set
Jira
Gmail
42 ノードBilly Christi
マルチモーダルAI
n8nノードの探索（可視化リファレンスライブラリ内）
n8nノードを可視化リファレンスライブラリで探索
If
Ftp
Set
+
If
Ftp
Set
113 ノードI versus AI
その他
コンテンツジェネレーター v3
AI驱动ブログ自動化：使用GPT-4生成并公開SEO記事至WordPressとTwitter
If
Set
Code
+
If
Set
Code
144 ノードJay Emp0
コンテンツ作成
APIアーキテクチャ抽出ツール
APIアーキテクチャ抽出器
If
Set
Code
+
If
Set
Code
88 ノードPolina Medvedieva
エンジニアリング
