338 測定値

LLM を使用して CSV ファイルをグラフに変換する: ステップバイステップガイド

に Neo4j41m2024/10/29

長すぎる; 読むには

LLM を使用して CSV ファイルをグラフ構造に変換し、反復的なプロンプトベースのアプローチで Neo4j のデータモデリングを改善する方法を学びます。

featured image - LLM を使用して CSV ファイルをグラフに変換する: ステップバイステップガイド

フラット CSV ファイルからグラフを作成しようとすると、LLM はどのように機能しますか?

私の仕事の大部分は、Neo4j でのユーザーエクスペリエンスを向上させることです。多くの場合、データを Neo4j に取り込み、効率的にモデル化することは、特に初期の段階では、ユーザーにとって大きな課題となります。初期のデータモデルは重要であり、熟考が必要ですが、データサイズやユーザー数の増加に応じて、簡単にリファクタリングしてパフォーマンスを向上させることができます。

そこで、自分への挑戦として、LLM が初期データモデルの作成に役立つかどうか試してみようと思いました。少なくとも、物事がどのようにつながっているかを示し、ユーザーが他の人に見せることができる簡単な結果を提供します。

直感的に、データモデリングは反復的なプロセスであり、特定の LLM は大量のデータによって簡単に混乱する可能性があることはわかっています。そのため、これはLangGraph を使用してデータを循環的に処理する良い機会となりました。

それを実現したプロンプトについて詳しく見てみましょう。

グラフモデリングの基礎

GraphAcademy のグラフデータモデリングの基礎コースでは、グラフ内のデータモデリングの基礎について説明していますが、最初のステップとして、次の経験則を使用します。

名詞はラベルになり、ノードが表すものを説明します。
動詞は関係性のタイプになり、物事がどのようにつながっているかを説明します。
その他すべてはプロパティになります (特に副詞) — あなたには名前があり、灰色の車を運転しているかもしれません。

動詞もノードになることができます。ある人が商品を注文したことを知るのは嬉しいかもしれませんが、その基本モデルでは、商品がいつどこで注文されたかを知ることはできません。この場合、注文はモデル内の新しいノードになります。

これを要約すれば、グラフデータモデリングへのゼロショットアプローチを作成するためのプロンプトになるはずです。

反復的なアプローチ

数か月前にこれを簡単に試してみたところ、使用していたモデルは大きなスキーマを扱うときに簡単に気が散ってしまい、プロンプトがすぐに LLM のトークン制限に達することがわかりました。

今回は、キーを 1 つずつ取得する反復的なアプローチを試してみようと思いました。LLM は一度に 1 つの項目のみを考慮する必要があるため、これにより集中力が散漫になるのを防ぐことができます。

最終的なアプローチでは、次の手順が使用されました。

CSV ファイルを Pandas データフレームに読み込みます。
CSV 内の各列を分析し、JSON スキーマに大まかに基づいたデータモデルに追加します。
各エンティティの不足している一意の ID を識別して追加します。
データモデルの正確性を確認します。
ノードとリレーションシップをインポートするための Cypher ステートメントを生成します。
インポートステートメントの基礎となる一意の制約を生成します。
制約を作成し、インポートを実行します。

データ

興味深いデータセットを探すために Kaggle をざっと見てみました。目立ったデータセットはSpotify Most Streamed Songsでした。

 import pandas as pd csv_file = '/Users/adam/projects/datamodeller/data/spotify/spotify-most-streamed-songs.csv' df = pd.read_csv(csv_file) df.head() track_name artist(s)_name artist_count released_year released_month released_day in_spotify_playlists in_spotify_charts streams in_apple_playlists … key mode danceability_% valence_% energy_% acousticness_% instrumentalness_% liveness_% speechiness_% cover_url 0 Seven (feat. Latto) (Explicit Ver.) Latto, Jung Kook 2 2023 7 14 553 147 141381703 43 … B Major 80 89 83 31 0 8 4 Not Found 1 LALA Myke Towers 1 2023 3 23 1474 48 133716286 48 … C# Major 71 61 74 7 0 10 4 https://i.scdn.co/image/ab67616d0000b2730656d5… 2 vampire Olivia Rodrigo 1 2023 6 30 1397 113 140003974 94 … F Major 51 32 53 17 0 31 6 https://i.scdn.co/image/ab67616d0000b273e85259… 3 Cruel Summer Taylor Swift 1 2019 8 23 7858 100 800840817 116 … A Major 55 58 72 11 0 11 15 https://i.scdn.co/image/ab67616d0000b273e787cf… 4 WHERE SHE GOES Bad Bunny 1 2023 5 18 3133 50 303236322 84 … A Minor 65 23 80 14 63 11 6 https://i.scdn.co/image/ab67616d0000b273ab5c9c…

5行×25列

比較的シンプルですが、トラックとアーティストの間に関係性があるはずだということはすぐにわかります。

また、列名とアーティストが artist(s)_name 列内でコンマ区切りの値であるという点において、克服すべきデータクリーン度の課題もあります。

LLMを選択する

本当は地元の LLM を使いたかったのですが、Llama 3 では不十分であることが早い段階でわかりました。疑問がある場合は、OpenAI に頼ってください。

 from langchain_core.prompts import PromptTemplate from langchain_core.pydantic_v1 import BaseModel, Field from typing import List from langchain_core.output_parsers import JsonOutputParser from langchain_openai import ChatOpenAI llm = ChatOpenAI(model="gpt-4o")

データモデルの作成

データモデリングプロンプトを作成するために、簡略化されたモデリング手順を使用しました。一貫した出力を得るには、プロンプトを数回設計する必要がありました。

ゼロショットの例は比較的うまく機能しましたが、出力に一貫性がないことがわかりました。JSON 出力を保持するための構造化された出力を定義すると、非常に役立ちました。

 class JSONSchemaSpecification(BaseModel): notes: str = Field(description="Any notes or comments about the schema") jsonschema: str = Field(description="A JSON array of JSON schema specifications that describe the entities in the data model")

少数ショットの例の出力

JSON 自体も一貫性がなかったため、最終的には映画の推奨データセットに基づいてスキーマを定義することになりました。

出力例:

 example_output = [ dict( title="Person", type="object", description="Node", properties=[ dict(name="name", column_name="person_name", type="string", description="The name of the person", examples=["Tom Hanks"]), dict(name="date_of_birth", column_name="person_dob", type="date", description="The date of birth for the person", examples=["1987-06-05"]), dict(name="id", column_name="person_name, date_of_birth", type="string", description="The ID is a combination of name and date of birth to ensure uniqueness", examples=["tom-hanks-1987-06-05"]), ], ), dict( title="Director", type="object", description="Node", properties=[ dict(name="name", column_name="director_names", type="string", description="The name of the directors. Split values in column by a comma", examples=["Francis Ford Coppola"]), ], ), dict( title="Movie", type="object", description="Node", properties=[ dict(name="title", column_name="title", type="string", description="The title of the movie", examples=["Toy Story"]), dict(name="released", column_name="released", type="integer", description="The year the movie was released", examples=["1990"]), ], ), dict( title="ACTED_IN", type="object", description="Relationship", properties=[ dict(name="_from", column_name="od", type="string", description="Person found by the `id`. The ID is a combination of name and date of birth to ensure uniqueness", examples=["Person"]), dict(name="_to", column_name="title", type="string", description="The movie title", examples=["Movie"]), dict(name="roles", type="string", column_name="person_roles", description="The roles the person played in the movie", examples=["Woody"]), ], ), dict( title="DIRECTED", type="object", description="Relationship", properties=[ dict(name="_from", type="string", column_name="director_names", description="Director names are comma separated", examples=["Director"]), dict(name="_to", type="string", column_name="title", description="The label of the node this relationship ends at", examples=["Movie"]), ], ), ]

LLM がインポートスクリプトを生成できるように、厳密な JSON スキーマから逸脱し、column_name フィールドを出力に追加する必要がありました。説明の例を提供することもこの点で役立ちました。そうしないと、MATCH 句で使用されるプロパティが不一致になります。

チェーン

最後のプロンプトは次のとおりです。

 model_prompt = PromptTemplate.from_template(""" You are an expert Graph Database administrator. Your task is to design a data model based on the information provided from an existing data source. You must decide where the following column fits in with the existing data model. Consider: * Does the column represent an entity, for example a Person, Place, or Movie? If so, this should be a node in its own right. * Does the column represent a relationship between two entities? If so, this should be a relationship between two nodes. * Does the column represent an attribute of an entity or relationship? If so, this should be a property of a node or relationship. * Does the column represent a shared attribute that could be interesting to query through to find similar nodes, for example a Genre? If so, this should be a node in its own right. ## Instructions for Nodes * Node labels are generally nouns, for example Person, Place, or Movie * Node titles should be in UpperCamelCase ## Instructions for Relationships * Relationshops are generally verbs, for example ACTED_IN, DIRECTED, or PURCHASED * Examples of good relationships are (:Person)-[:ACTED_IN]->(:Movie) or (:Person)-[:PURCHASED]->(:Product) * Relationships should be in UPPER_SNAKE_CASE * Provide any specific instructions for the field in the description. For example, does the field contain a list of comma separated values or a single value? ## Instructions for Properties * Relationships should be in lowerPascalCase * Prefer the shorter name where possible, for example "person_id" and "personId" should simply be "id" * If you are changing the property name from the original field name, mention the column name in the description * Do not include examples for integer or date fields * Always include instructions on data preparation for the field. Does it need to be cast as a string or split into multiple fields on a delimiting value? * Property keys should be letters only, no numbers or special characters. ## Important! Consider the examples provided. Does any data preparation need to be done to ensure the data is in the correct format? You must include any information about data preparation in the description. ## Example Output Here is an example of a good output: {example_output} ## New Data: Key: {key} Data Type: {type} Example Values: {examples} ## Existing Data Model Here is the existing data model: {existing_model} ## Keep Existing Data Model Apply your changes to the existing data model but never remove any existing definitions. """, partial_variables=dict(example_output=dumps(example_output))) model_chain = model_prompt | llm.with_structured_output(JSONSchemaSpecification)

チェーンの実行

モデルを反復的に更新するために、データフレーム内のキーを反復処理し、各キー、そのデータ型、および最初の 5 つの一意の値をプロンプトに渡しました。

 from json_repair import dumps, loads existing_model = {} for i, key in enumerate(df): print("\n", i, key) print("----------------") try: res = try_chain(model_chain, dict( existing_model=dumps(existing_model), key=key, type=df[key].dtype, examples=dumps(df[key].unique()[:5].tolist()) )) print(res.notes) existing_model = loads(res.jsonschema) print([n['title'] for n in existing_model]) except Exception as e: print(e) pass existing_model

コンソール出力:

 0 track_name ---------------- Adding 'track_name' to an existing data model. This represents a music track entity. ['Track'] 1 artist(s)_name ---------------- Adding a new column 'artist(s)_name' to the existing data model. This column represents multiple artists associated with tracks and should be modeled as a new node 'Artist' and a relationship 'PERFORMED_BY' from 'Track' to 'Artist'. ['Track', 'Artist', 'PERFORMED_BY'] 2 artist_count ---------------- Added artist_count as a property of Track node. This property indicates the number of artists performing in the track. ['Track', 'Artist', 'PERFORMED_BY'] 3 released_year ---------------- Add the released_year column to the existing data model as a property of the Track node. ['Track', 'Artist', 'PERFORMED_BY'] 4 released_month ---------------- Adding the 'released_month' column to the existing data model, considering it as an attribute of the Track node. ['Track', 'Artist', 'PERFORMED_BY'] 5 released_day ---------------- Added a new property 'released_day' to the 'Track' node to capture the day of the month a track was released. ['Track', 'Artist', 'PERFORMED_BY'] 6 in_spotify_playlists ---------------- Adding the new column 'in_spotify_playlists' to the existing data model as a property of the 'Track' node. ['Track', 'Artist', 'PERFORMED_BY'] 7 in_spotify_charts ---------------- Adding the 'in_spotify_charts' column to the existing data model as a property of the Track node. ['Track', 'Artist', 'PERFORMED_BY'] 8 streams ---------------- Adding a new column 'streams' to the existing data model, representing the number of streams for a track. ['Track', 'Artist', 'PERFORMED_BY'] 9 in_apple_playlists ---------------- Adding new column 'in_apple_playlists' to the existing data model ['Track', 'Artist', 'PERFORMED_BY'] 10 in_apple_charts ---------------- Adding 'in_apple_charts' as a property to the 'Track' node, representing the number of times the track appeared in the Apple charts. ['Track', 'Artist', 'PERFORMED_BY'] 11 in_deezer_playlists ---------------- Add 'in_deezer_playlists' to the existing data model for a music track database. ['Track', 'Artist', 'PERFORMED_BY'] 12 in_deezer_charts ---------------- Adding a new property 'inDeezerCharts' to the existing 'Track' node to represent the number of times the track appeared in Deezer charts. ['Track', 'Artist', 'PERFORMED_BY'] 13 in_shazam_charts ---------------- Adding new data 'in_shazam_charts' to the existing data model. This appears to be an attribute of the 'Track' node, indicating the number of times a track appeared in the Shazam charts. ['Track', 'Artist', 'PERFORMED_BY'] 14 bpm ---------------- Added bpm column as a property to the Track node as it represents a characteristic of the track. ['Track', 'Artist', 'PERFORMED_BY'] 15 key ---------------- Adding the 'key' column to the existing data model. The 'key' represents the musical key of a track, which is a shared attribute that can be interesting to query through to find similar tracks. ['Track', 'Artist', 'PERFORMED_BY'] 16 mode ---------------- Adding 'mode' to the existing data model. It represents a musical characteristic of a track, which is best captured as an attribute of the Track node. ['Track', 'Artist', 'PERFORMED_BY'] 17 danceability_% ---------------- Added 'danceability_%' to the existing data model as a property of the Track node. The field represents the danceability percentage of the track. ['Track', 'Artist', 'PERFORMED_BY'] 18 valence_% ---------------- Adding the valence percentage column to the existing data model as a property of the Track node. ['Track', 'Artist', 'PERFORMED_BY'] 19 energy_% ---------------- Integration of the new column 'energy_%' into the existing data model. This column represents an attribute of the Track entity and should be added as a property of the Track node. ['Track', 'Artist', 'PERFORMED_BY'] 20 acousticness_% ---------------- Adding acousticness_% to the existing data model as a property of the Track node. ['Track', 'Artist', 'PERFORMED_BY'] 21 instrumentalness_% ---------------- Adding the new column 'instrumentalness_%' to the existing Track node as it represents an attribute of the Track entity. ['Track', 'Artist', 'PERFORMED_BY'] 22 liveness_% ---------------- Adding the new column 'liveness_%' to the existing data model as an attribute of the Track node ['Track', 'Artist', 'PERFORMED_BY'] 23 speechiness_% ---------------- Adding the new column 'speechiness_%' to the existing data model as a property of the 'Track' node. ['Track', 'Artist', 'PERFORMED_BY'] 24 cover_url ---------------- Adding a new property 'cover_url' to the existing 'Track' node. This property represents the URL of the track's cover image. ['Track', 'Artist', 'PERFORMED_BY']

ユースケースを処理するためにプロンプトを少し調整した後、満足のいくモデルが完成しました。LLM は、データセットが Track、Artist、および 2 つを接続する PERFORMED_BY 関係で構成されていることを判別しました。

 [ { "title": "Track", "type": "object", "description": "Node", "properties": [ { "name": "name", "column_name": "track_name", "type": "string", "description": "The name of the track", "examples": [ "Seven (feat. Latto) (Explicit Ver.)", "LALA", "vampire", "Cruel Summer", "WHERE SHE GOES", ], }, { "name": "artist_count", "column_name": "artist_count", "type": "integer", "description": "The number of artists performing in the track", "examples": [2, 1, 3, 8, 4], }, { "name": "released_year", "column_name": "released_year", "type": "integer", "description": "The year the track was released", "examples": [2023, 2019, 2022, 2013, 2014], }, { "name": "released_month", "column_name": "released_month", "type": "integer", "description": "The month the track was released", "examples": [7, 3, 6, 8, 5], }, { "name": "released_day", "column_name": "released_day", "type": "integer", "description": "The day of the month the track was released", "examples": [14, 23, 30, 18, 1], }, { "name": "inSpotifyPlaylists", "column_name": "in_spotify_playlists", "type": "integer", "description": "The number of Spotify playlists the track is in. Cast the value as an integer.", "examples": [553, 1474, 1397, 7858, 3133], }, { "name": "inSpotifyCharts", "column_name": "in_spotify_charts", "type": "integer", "description": "The number of times the track appeared in the Spotify charts. Cast the value as an integer.", "examples": [147, 48, 113, 100, 50], }, { "name": "streams", "column_name": "streams", "type": "array", "description": "The list of stream IDs for the track. Maintain the array format.", "examples": [ "141381703", "133716286", "140003974", "800840817", "303236322", ], }, { "name": "inApplePlaylists", "column_name": "in_apple_playlists", "type": "integer", "description": "The number of Apple playlists the track is in. Cast the value as an integer.", "examples": [43, 48, 94, 116, 84], }, { "name": "inAppleCharts", "column_name": "in_apple_charts", "type": "integer", "description": "The number of times the track appeared in the Apple charts. Cast the value as an integer.", "examples": [263, 126, 207, 133, 213], }, { "name": "inDeezerPlaylists", "column_name": "in_deezer_playlists", "type": "array", "description": "The list of Deezer playlist IDs the track is in. Maintain the array format.", "examples": ["45", "58", "91", "125", "87"], }, { "name": "inDeezerCharts", "column_name": "in_deezer_charts", "type": "integer", "description": "The number of times the track appeared in the Deezer charts. Cast the value as an integer.", "examples": [10, 14, 12, 15, 17], }, { "name": "inShazamCharts", "column_name": "in_shazam_charts", "type": "array", "description": "The list of Shazam chart IDs the track is in. Maintain the array format.", "examples": ["826", "382", "949", "548", "425"], }, { "name": "bpm", "column_name": "bpm", "type": "integer", "description": "The beats per minute of the track. Cast the value as an integer.", "examples": [125, 92, 138, 170, 144], }, { "name": "key", "column_name": "key", "type": "string", "description": "The musical key of the track. Cast the value as a string.", "examples": ["B", "C#", "F", "A", "D"], }, { "name": "mode", "column_name": "mode", "type": "string", "description": "The mode of the track (eg, Major, Minor). Cast the value as a string.", "examples": ["Major", "Minor"], }, { "name": "danceability", "column_name": "danceability_%", "type": "integer", "description": "The danceability percentage of the track. Cast the value as an integer.", "examples": [80, 71, 51, 55, 65], }, { "name": "valence", "column_name": "valence_%", "type": "integer", "description": "The valence percentage of the track. Cast the value as an integer.", "examples": [89, 61, 32, 58, 23], }, { "name": "energy", "column_name": "energy_%", "type": "integer", "description": "The energy percentage of the track. Cast the value as an integer.", "examples": [83, 74, 53, 72, 80], }, { "name": "acousticness", "column_name": "acousticness_%", "type": "integer", "description": "The acousticness percentage of the track. Cast the value as an integer.", "examples": [31, 7, 17, 11, 14], }, { "name": "instrumentalness", "column_name": "instrumentalness_%", "type": "integer", "description": "The instrumentalness percentage of the track. Cast the value as an integer.", "examples": [0, 63, 17, 2, 19], }, { "name": "liveness", "column_name": "liveness_%", "type": "integer", "description": "The liveness percentage of the track. Cast the value as an integer.", "examples": [8, 10, 31, 11, 28], }, { "name": "speechiness", "column_name": "speechiness_%", "type": "integer", "description": "The speechiness percentage of the track. Cast the value as an integer.", "examples": [4, 6, 15, 24, 3], }, { "name": "coverUrl", "column_name": "cover_url", "type": "string", "description": "The URL of the track's cover image. If the value is 'Not Found', it should be cast as an empty string.", "examples": [ "https://i.scdn.co/image/ab67616d0000b2730656d5ce813ca3cc4b677e05", "https://i.scdn.co/image/ab67616d0000b273e85259a1cae29a8d91f2093d", ], }, ], }, { "title": "Artist", "type": "object", "description": "Node", "properties": [ { "name": "name", "column_name": "artist(s)_name", "type": "string", "description": "The name of the artist. Split values in column by a comma", "examples": [ "Latto", "Jung Kook", "Myke Towers", "Olivia Rodrigo", "Taylor Swift", "Bad Bunny", ], } ], }, { "title": "PERFORMED_BY", "type": "object", "description": "Relationship", "properties": [ { "name": "_from", "type": "string", "description": "The label of the node this relationship starts at", "examples": ["Track"], }, { "name": "_to", "type": "string", "description": "The label of the node this relationship ends at", "examples": ["Artist"], }, ], }, ] [ { "title": "Track", "type": "object", "description": "Node", "properties": [ { "name": "name", "column_name": "track_name", "type": "string", "description": "The name of the track", "examples": [ "Seven (feat. Latto) (Explicit Ver.)", "LALA", "vampire", "Cruel Summer", "WHERE SHE GOES", ], }, { "name": "artist_count", "column_name": "artist_count", "type": "integer", "description": "The number of artists performing in the track", "examples": [2, 1, 3, 8, 4], }, { "name": "released_year", "column_name": "released_year", "type": "integer", "description": "The year the track was released", "examples": [2023, 2019, 2022, 2013, 2014], }, { "name": "released_month", "column_name": "released_month", "type": "integer", "description": "The month the track was released", "examples": [7, 3, 6, 8, 5], }, { "name": "released_day", "column_name": "released_day", "type": "integer", "description": "The day of the month the track was released", "examples": [14, 23, 30, 18, 1], }, { "name": "inSpotifyPlaylists", "column_name": "in_spotify_playlists", "type": "integer", "description": "The number of Spotify playlists the track is in. Cast the value as an integer.", "examples": [553, 1474, 1397, 7858, 3133], }, { "name": "inSpotifyCharts", "column_name": "in_spotify_charts", "type": "integer", "description": "The number of times the track appeared in the Spotify charts. Cast the value as an integer.", "examples": [147, 48, 113, 100, 50], }, { "name": "streams", "column_name": "streams", "type": "array", "description": "The list of stream IDs for the track. Maintain the array format.", "examples": [ "141381703", "133716286", "140003974", "800840817", "303236322", ], }, { "name": "inApplePlaylists", "column_name": "in_apple_playlists", "type": "integer", "description": "The number of Apple playlists the track is in. Cast the value as an integer.", "examples": [43, 48, 94, 116, 84], }, { "name": "inAppleCharts", "column_name": "in_apple_charts", "type": "integer", "description": "The number of times the track appeared in the Apple charts. Cast the value as an integer.", "examples": [263, 126, 207, 133, 213], }, { "name": "inDeezerPlaylists", "column_name": "in_deezer_playlists", "type": "array", "description": "The list of Deezer playlist IDs the track is in. Maintain the array format.", "examples": ["45", "58", "91", "125", "87"], }, { "name": "inDeezerCharts", "column_name": "in_deezer_charts", "type": "integer", "description": "The number of times the track appeared in the Deezer charts. Cast the value as an integer.", "examples": [10, 14, 12, 15, 17], }, { "name": "inShazamCharts", "column_name": "in_shazam_charts", "type": "array", "description": "The list of Shazam chart IDs the track is in. Maintain the array format.", "examples": ["826", "382", "949", "548", "425"], }, { "name": "bpm", "column_name": "bpm", "type": "integer", "description": "The beats per minute of the track. Cast the value as an integer.", "examples": [125, 92, 138, 170, 144], }, { "name": "key", "column_name": "key", "type": "string", "description": "The musical key of the track. Cast the value as a string.", "examples": ["B", "C#", "F", "A", "D"], }, { "name": "mode", "column_name": "mode", "type": "string", "description": "The mode of the track (eg, Major, Minor). Cast the value as a string.", "examples": ["Major", "Minor"], }, { "name": "danceability", "column_name": "danceability_%", "type": "integer", "description": "The danceability percentage of the track. Cast the value as an integer.", "examples": [80, 71, 51, 55, 65], }, { "name": "valence", "column_name": "valence_%", "type": "integer", "description": "The valence percentage of the track. Cast the value as an integer.", "examples": [89, 61, 32, 58, 23], }, { "name": "energy", "column_name": "energy_%", "type": "integer", "description": "The energy percentage of the track. Cast the value as an integer.", "examples": [83, 74, 53, 72, 80], }, { "name": "acousticness", "column_name": "acousticness_%", "type": "integer", "description": "The acousticness percentage of the track. Cast the value as an integer.", "examples": [31, 7, 17, 11, 14], }, { "name": "instrumentalness", "column_name": "instrumentalness_%", "type": "integer", "description": "The instrumentalness percentage of the track. Cast the value as an integer.", "examples": [0, 63, 17, 2, 19], }, { "name": "liveness", "column_name": "liveness_%", "type": "integer", "description": "The liveness percentage of the track. Cast the value as an integer.", "examples": [8, 10, 31, 11, 28], }, { "name": "speechiness", "column_name": "speechiness_%", "type": "integer", "description": "The speechiness percentage of the track. Cast the value as an integer.", "examples": [4, 6, 15, 24, 3], }, { "name": "coverUrl", "column_name": "cover_url", "type": "string", "description": "The URL of the track's cover image. If the value is 'Not Found', it should be cast as an empty string.", "examples": [ "https://i.scdn.co/image/ab67616d0000b2730656d5ce813ca3cc4b677e05", "https://i.scdn.co/image/ab67616d0000b273e85259a1cae29a8d91f2093d", ], }, ], }, { "title": "Artist", "type": "object", "description": "Node", "properties": [ { "name": "name", "column_name": "artist(s)_name", "type": "string", "description": "The name of the artist. Split values in column by a comma", "examples": [ "Latto", "Jung Kook", "Myke Towers", "Olivia Rodrigo", "Taylor Swift", "Bad Bunny", ], } ], }, { "title": "PERFORMED_BY", "type": "object", "description": "Relationship", "properties": [ { "name": "_from", "type": "string", "description": "The label of the node this relationship starts at", "examples": ["Track"], }, { "name": "_to", "type": "string", "description": "The label of the node this relationship ends at", "examples": ["Artist"], }, ], }, ]

一意の識別子の追加

スキーマに一意の識別子が含まれていないことに気付きました。これは、関係をインポートするときに問題になる可能性があります。異なるアーティストが同じ名前の曲をリリースしたり、2 人のアーティストが同じ名前を持つのは当然です。

このため、より大きなデータセット内でトラックを区別できるように、トラックの識別子を作成することが重要でした。

 # Add primary key/unique identifiers uid_prompt = PromptTemplate.from_template(""" You are a graph database expert reviewing a single entity from a data model generated by a colleague. You want to ensure that all of the nodes imported into the database are unique. ## Example A schema contains Actors with a number of properties including name, date of birth. Two actors may have the same name then add a new compound property combining the name and date of birth. If combining values, include the instruction to convert the value to slug case. Call the new property 'id'. If you have identified a new property, add it to the list of properties leaving the rest intact. Include in the description the fields that are to be concatenated. ## Example Output Here is an example of a good output: {example_output} ## Current Entity Schema {entity} """, partial_variables=dict(example_output=dumps(example_output))) uid_chain = uid_prompt | llm.with_structured_output(JSONSchemaSpecification)

この手順は実際にはノードにのみ必要なので、スキーマからノードを抽出し、それぞれに対してチェーンを実行してから、更新された定義と関係を結合しました。

 # extract nodes and relationships nodes = [n for n in existing_model if "node" in n["description"].lower()] rels = [n for n in existing_model if "node" not in n["description"].lower()] # generate a unique id for nodes with_uids = [] for entity in nodes: res = uid_chain.invoke(dict(entity=dumps(entity))) json = loads(res.jsonschema) with_uids = with_uids + json if type(json) == list else with_uids + [json] # combine nodes and relationships with_uids = with_uids + rels

データモデルのレビュー

健全性のために、モデルの最適化をチェックする価値はあります。model_prompt は名詞と動詞を適切に識別しましたが、より複雑なモデルでした。

1 回の反復では、*_playlists 列と _charts 列を ID として扱い、ストリームノードと IN_PLAYLIST 関係を作成しようとしました。これは、カンマによる書式設定を含むカウントが 1,000 を超えたためであると考えられます (例: 1,001)。

素晴らしいアイデアですが、ちょっと賢すぎるかもしれません。しかし、これは、データ構造を理解する人間が関与することの重要性を示しています。

 # Add primary key/unique identifiers review_prompt = PromptTemplate.from_template(""" You are a graph database expert reviewing a data model generated by a colleague. Your task is to review the data model and ensure that it is fit for purpose. Check for: ## Check for nested objects Remember that Neo4j cannot store arrays of objects or nested objects. These must be converted into into separate nodes with relationships between them. You must include the new node and a reference to the relationship to the output schema. ## Check for Entities in properties If there is a property that represents an array of IDs, a new node should be created for that entity. You must include the new node and a reference to the relationship to the output schema. # Keep Instructions Ensure that the instructions for the nodes, relationships, and properties are clear and concise. You may improve them but the detail must not be removed in any circumstances. ## Current Entity Schema {entity} """) review_chain = review_prompt | llm.with_structured_output(JSONSchemaSpecification) review_nodes = [n for n in with_uids if "node" in n["description"].lower() ] review_rels = [n for n in with_uids if "node" not in n["description"].lower() ] reviewed = [] for entity in review_nodes: res = review_chain.invoke(dict(entity=dumps(entity))) json = loads(res.jsonschema) reviewed = reviewed + json # add relationships back in reviewed = reviewed + review_rels len(reviewed) reviewed = with_uids

実際のシナリオでは、これを数回実行して、データモデルを反復的に改善する必要があります。最大制限を設定し、そのポイントまで反復すると、データモデルオブジェクトは変更されなくなります。

インポートステートメントを生成する

この時点で、スキーマは十分に堅牢であり、LLM がインポートスクリプトのセットを生成できるようにできるだけ多くの情報が含まれている必要があります。

Neo4j データのインポートに関する推奨事項に従って、集中的な操作とロックを回避するために、ファイルは複数回処理され、そのたびに単一のノードまたはリレーションシップがインポートされる必要があります。

 import_prompt = PromptTemplate.from_template(""" Based on the data model, write a Cypher statement to import the following data from a CSV file into Neo4j. Do not use LOAD CSV as this data will be imported using the Neo4j Python Driver, use UNWIND on the $rows parameter instead. You are writing a multi-pass import process, so concentrate on the entity mentioned. When importing data, you must use the following guidelines: * follow the instructions in the description when identifying primary keys. * Use the instructions in the description to determine the format of properties when a finding. * When combining fields into an ID, use the apoc.text.slug function to convert any text to slug case and toLower to convert the string to lowercase - apoc.text.slug(toLower(row.`name`)) * If you split a property, convert it to a string and use the trim function to remove any whitespace - trim(toString(row.`name`)) * When combining properties, wrap each property in the coalesce function so the property is not null if one of the values is not set - coalesce(row.`id`, '') + '--'+ coalsece(row.`title`) * Use the `column_name` field to map the CSV column to the property in the data model. * Wrap all column names from the CSV in backticks - for example row.`column_name`. * When you merge nodes, merge on the unique identifier and nothing else. All other properties should be set using `SET`. * Do not use apoc.periodic.iterate, the files will be batched in the application. Data Model: {data_model} Current Entity: {entity} """)

このチェーンでは、前のステップとは異なる出力オブジェクトが必要です。この場合、 cypher メンバーが最も重要ですが、Chain of Thought を促進するために chain_of_thought キーも含めたいと考えました。

 class CypherOutputSpecification(BaseModel): chain_of_thought: str = Field(description="Any reasoning used to write the Cypher statement") cypher: str = Field(description="The Cypher statement to import the data") notes: Optional[str] = Field(description="Any notes or closing remarks about the Cypher statement") import_chain = import_prompt | llm.with_structured_output(CypherOutputSpecification)

次に、同じプロセスを適用して、レビューされた各定義を反復処理し、Cypher を生成します。

 import_cypher = [] for n in reviewed: print('\n\n------', n['title']) res = import_chain.invoke(dict( data_model=dumps(reviewed), entity=n )) import_cypher.append(( res.cypher )) print(res.cypher)

コンソール出力:

 ------ Track UNWIND $rows AS row MERGE (t:Track {id: apoc.text.slug(toLower(coalesce(row.`track_name`, '') + '-' + coalesce(row.`released_year`, '')))}) SET t.name = trim(toString(row.`track_name`)), t.artist_count = toInteger(row.`artist_count`), t.released_year = toInteger(row.`released_year`), t.released_month = toInteger(row.`released_month`), t.released_day = toInteger(row.`released_day`), t.inSpotifyPlaylists = toInteger(row.`in_spotify_playlists`), t.inSpotifyCharts = toInteger(row.`in_spotify_charts`), t.streams = row.`streams`, t.inApplePlaylists = toInteger(row.`in_apple_playlists`), t.inAppleCharts = toInteger(row.`in_apple_charts`), t.inDeezerPlaylists = row.`in_deezer_playlists`, t.inDeezerCharts = toInteger(row.`in_deezer_charts`), t.inShazamCharts = row.`in_shazam_charts`, t.bpm = toInteger(row.`bpm`), t.key = trim(toString(row.`key`)), t.mode = trim(toString(row.`mode`)), t.danceability = toInteger(row.`danceability_%`), t.valence = toInteger(row.`valence_%`), t.energy = toInteger(row.`energy_%`), t.acousticness = toInteger(row.`acousticness_%`), t.instrumentalness = toInteger(row.`instrumentalness_%`), t.liveness = toInteger(row.`liveness_%`), t.speechiness = toInteger(row.`speechiness_%`), t.coverUrl = CASE row.`cover_url` WHEN 'Not Found' THEN '' ELSE trim(toString(row.`cover_url`)) END ------ Artist UNWIND $rows AS row WITH row, split(row.`artist(s)_name`, ',') AS artistNames UNWIND artistNames AS artistName MERGE (a:Artist {id: apoc.text.slug(toLower(trim(artistName)))}) SET a.name = trim(artistName) ------ PERFORMED_BY UNWIND $rows AS row UNWIND split(row.`artist(s)_name`, ',') AS artist_name MERGE (t:Track {id: apoc.text.slug(toLower(row.`track_name`)) + '-' + trim(toString(row.`released_year`))}) MERGE (a:Artist {id: apoc.text.slug(toLower(trim(artist_name)))}) MERGE (t)-[:PERFORMED_BY]->(a)

このプロンプトでは、一貫した結果を得るために、ある程度のエンジニアリングが必要でした。

Cypher には、複数のフィールドが定義された MERGE ステートメントが含まれることがありますが、これは最適とは言えません。いずれかの列が null の場合、インポート全体が失敗します。
場合によっては、結果にapoc.period.iterateが含まれますが、これはもう必要ありません。Python ドライバーで実行できるコードが必要でした。
リレーションシップを作成するときは、指定された列名を使用する必要があることを繰り返し強調する必要がありました。
LLM は、関係の両端のノードで一意の識別子を使用する場合、指示に従わなかったため、説明の指示に従うようにするには、数回の試行が必要でした。このプロンプトと model_prompt の間で何度かやり取りがありました。
特殊文字 (例: energy_%) を含む列名にはバックティックが必要でした。

これを 2 つのプロンプト (ノード用と関係用) に分割することも有益です。ただし、これは別の日に行う作業です。

ユニーク制約を作成する

次に、インポートスクリプトを基礎として使用して、データベースに一意の制約を作成できます。

 constraint_prompt = PromptTemplate.from_template(""" You are an expert graph database administrator. Use the following Cypher statement to write a Cypher statement to create unique constraints on any properties used in a MERGE statement. The correct syntax for a unique constraint is: CREATE CONSTRAINT movie_title_id IF NOT EXISTS FOR (m:Movie) REQUIRE m.title IS UNIQUE; Cypher: {cypher} """) constraint_chain = constraint_prompt | llm.with_structured_output(CypherOutputSpecification) constraint_queries = [] for statement in import_cypher: res = constraint_chain.invoke(dict(cypher=statement)) statements = res.cypher.split(";") for cypher in statements: constraint_queries.append(cypher)

コンソール出力:

 CREATE CONSTRAINT track_id_unique IF NOT EXISTS FOR (t:Track) REQUIRE t.id IS UNIQUE CREATE CONSTRAINT stream_id IF NOT EXISTS FOR (s:Stream) REQUIRE s.id IS UNIQUE CREATE CONSTRAINT playlist_id IF NOT EXISTS FOR (p:Playlist) REQUIRE p.id IS UNIQUE CREATE CONSTRAINT chart_id IF NOT EXISTS FOR (c:Chart) REQUIRE c.id IS UNIQUE CREATE CONSTRAINT track_id_unique IF NOT EXISTS FOR (t:Track) REQUIRE t.id IS UNIQUE CREATE CONSTRAINT stream_id_unique IF NOT EXISTS FOR (s:Stream) REQUIRE s.id IS UNIQUE CREATE CONSTRAINT track_id_unique IF NOT EXISTS FOR (t:Track) REQUIRE t.id IS UNIQUE CREATE CONSTRAINT playlist_id_unique IF NOT EXISTS FOR (p:Playlist) REQUIRE p.id IS UNIQUE CREATE CONSTRAINT track_id_unique IF NOT EXISTS FOR (track:Track) REQUIRE track.id IS UNIQUE CREATE CONSTRAINT chart_id_unique IF NOT EXISTS FOR (chart:Chart) REQUIRE chart.id IS UNIQUE

このプロンプトは、インデックスと制約のステートメントを返すことがあるため、セミコロンで分割されます。

インポートを実行する

すべての準備が整ったら、Cypher ステートメントを実行します。

 from os import getenv from neo4j import GraphDatabase driver = GraphDatabase.driver( getenv("NEO4J_URI"), auth=( getenv("NEO4J_USERNAME"), getenv("NEO4J_PASSWORD") ) ) with driver.session() as session: # truncate the db session.run("MATCH (n) DETACH DELETE n") # create constraints for q in constraint_queries: if q.strip() != "": session.run(q) # import the data for q in import_cypher: if q.strip() != "": res = session.run(q, rows=rows).consume() print(q) print(res.counters)

データセットのQA

GraphCypherQAChain を使用したデータセットの QA がなければ、この投稿は完了しません。

 from langchain.chains import GraphCypherQAChain from langchain_community.graphs import Neo4jGraph graph = Neo4jGraph( url=getenv("NEO4J_URI"), username=getenv("NEO4J_USERNAME"), password=getenv("NEO4J_PASSWORD"), enhanced_schema=True ) qa = GraphCypherQAChain.from_llm( llm, graph=graph, allow_dangerous_requests=True, verbose=True )

最も人気のあるアーティスト

データベース内で最も人気のあるアーティストは誰ですか?

 qa.invoke({"query": "Who are the most popular artists?"}) > Entering new GraphCypherQAChain chain... Generated Cypher: cypher MATCH (:Track)-[:PERFORMED_BY]->(a:Artist) RETURN a.name, COUNT(*) AS popularity ORDER BY popularity DESC LIMIT 10 Full Context: [{'a.name': 'Bad Bunny', 'popularity': 40}, {'a.name': 'Taylor Swift', 'popularity': 38}, {'a.name': 'The Weeknd', 'popularity': 36}, {'a.name': 'SZA', 'popularity': 23}, {'a.name': 'Kendrick Lamar', 'popularity': 23}, {'a.name': 'Feid', 'popularity': 21}, {'a.name': 'Drake', 'popularity': 19}, {'a.name': 'Harry Styles', 'popularity': 17}, {'a.name': 'Peso Pluma', 'popularity': 16}, {'a.name': '21 Savage', 'popularity': 14}] > Finished chain. { "query": "Who are the most popular artists?", "result": "Bad Bunny, Taylor Swift, and The Weeknd are the most popular artists." }

LLM は、アーティストの人気を、全体のストリーム数ではなく、そのアーティストが参加したトラックの数で判断しているようです。

1分間の拍数

どのトラックのBPMが最も高いですか?

 qa.invoke({"query": "Which track has the highest BPM?"}) > Entering new GraphCypherQAChain chain... Generated Cypher: cypher MATCH (t:Track) RETURN t ORDER BY t.bpm DESC LIMIT 1 Full Context: [{'t': {'id': 'seven-feat-latto-explicit-ver--2023'}}] > Finished chain. { "query": "Which track has the highest BPM?", "result": "I don't know the answer." }

暗号生成プロンプトの改善

この場合、Cypher は正常に見え、正しい結果がプロンプトに含まれていましたが、gpt-4o は回答を解釈できませんでした。GraphCypherQAChain に渡される CYPHER_GENERATION_PROMPT には、列名をより詳細にするための追加の指示が必要であるようです。

Cypher ステートメントでは、ラベル名とプロパティ名を使用して、常に詳細な列名を使用します。たとえば、「name」の代わりに「person_name」を使用します。

カスタムプロンプト付きのGraphCypherQAChain:

 CYPHER_GENERATION_TEMPLATE = """Task:Generate Cypher statement to query a graph database. Instructions: Use only the provided relationship types and properties in the schema. Do not use any other relationship types or properties that are not provided. Schema: {schema} Note: Do not include any explanations or apologies in your responses. Do not respond to any questions that might ask anything else than for you to construct a Cypher statement. Do not include any text except the generated Cypher statement. Always use verbose column names in the Cypher statement using the label and property names. For example, use 'person_name' instead of 'name'. Include data from the immediate network around the node in the result to provide extra context. For example, include the Movie release year, a list of actors and their roles, or the director of a movie. When ordering by a property, add an `IS NOT NULL` check to ensure that only nodes with that property are returned. Examples: Here are a few examples of generated Cypher statements for particular questions: # How many people acted in Top Gun? MATCH (m:Movie {{name:"Top Gun"}}) RETURN COUNT { (m)<-[:ACTED_IN]-() } AS numberOfActors The question is: {question}""" CYPHER_GENERATION_PROMPT = PromptTemplate( input_variables=["schema", "question"], template=CYPHER_GENERATION_TEMPLATE ) qa = GraphCypherQAChain.from_llm( llm, graph=graph, allow_dangerous_requests=True, verbose=True, cypher_prompt=CYPHER_GENERATION_PROMPT, )

最も多くのアーティストが演奏したトラック

グラフは、タイプと方向別に関係の数を返すのに優れています。

 qa.invoke({"query": "Which tracks are performed by the most artists?"}) > Entering new GraphCypherQAChain chain... Generated Cypher: cypher MATCH (t:Track) WITH t, COUNT { (t)-[:PERFORMED_BY]->(:Artist) } as artist_count WHERE artist_count IS NOT NULL RETURN t.id AS track_id, t.name AS track_name, artist_count ORDER BY artist_count DESC Full Context: [{'track_id': 'los-del-espacio-2023', 'track_name': 'Los del Espacio', 'artist_count': 8}, {'track_id': 'se-le-ve-2021', 'track_name': 'Se Le Ve', 'artist_count': 8}, {'track_id': 'we-don-t-talk-about-bruno-2021', 'track_name': "We Don't Talk About Bruno", 'artist_count': 7}, {'track_id': 'cayï-ï-la-noche-feat-cruz-cafunï-ï-abhir-hathi-bejo-el-ima--2022', 'track_name': None, 'artist_count': 6}, {'track_id': 'jhoome-jo-pathaan-2022', 'track_name': 'Jhoome Jo Pathaan', 'artist_count': 6}, {'track_id': 'besharam-rang-from-pathaan--2022', 'track_name': None, 'artist_count': 6}, {'track_id': 'nobody-like-u-from-turning-red--2022', 'track_name': None, 'artist_count': 6}, {'track_id': 'ultra-solo-remix-2022', 'track_name': 'ULTRA SOLO REMIX', 'artist_count': 5}, {'track_id': 'angel-pt-1-feat-jimin-of-bts-jvke-muni-long--2023', 'track_name': None, 'artist_count': 5}, {'track_id': 'link-up-metro-boomin-don-toliver-wizkid-feat-beam-toian-spider-verse-remix-spider-man-across-the-spider-verse--2023', 'track_name': None, 'artist_count': 5}] > Finished chain. { "query": "Which tracks are performed by the most artists?", "result": "The tracks \"Los del Espacio\" and \"Se Le Ve\" are performed by the most artists, with each track having 8 artists." }

まとめ

CSV 分析とモデリングは最も時間のかかる部分です。生成には 5 分以上かかる場合があります。

コスト自体はかなり安かったです。8 時間の実験で何百ものリクエストを送信しましたが、結局 1 ドル程度しかかかりませんでした。

ここまで到達するまでには、数多くの課題がありました。

プロンプトが適切に表示されるまでには、数回の反復が必要でした。この問題は、モデルを微調整するか、少数のサンプルを提供することで解決できます。
GPT-4o からの JSON 応答は一貫性がない場合があります。 json-repairが推奨されましたが、これは LLM に独自の JSON 出力を検証させるよりも優れています。

このアプローチは、操作が順番に実行される LangGraph 実装でうまく機能し、LLM にモデルの構築と改良の機能を与えることがわかります。新しいモデルがリリースされると、微調整のメリットも得られる可能性があります。