Teams で RAG ボットを構築する

// create OpenAIEmbeddings instance
const model = new OpenAIEmbeddings({ ... endpoint, apikey, model, ... });

// create embeddings for the given inputs
const embeddings = await model.createEmbeddings(model, inputs);

// your own logic to process embeddings

# create OpenAIEmbeddings instance
model = OpenAIEmbeddings(OpenAIEmbeddingsOptions(api_key, model))

# create embeddings for the given inputs
embeddings = await model.create_embeddings(inputs)

# your own logic to process embeddings

• loader.ts: ソース入力としてのプレーン テキスト。

  import * as fs from "node:fs";
  
  export function loadTextFile(path: string): string {
    return fs.readFileSync(path, "utf-8");
  }
  

• splitter.ts: テキストをチャンクに分割し、重複します。

  
  // split words by delimiters.
  const delimiters = [" ", "\t", "\r", "\n"];
  
  export function split(content: string, length: number, overlap: number): Array<string> {
    const results = new Array<string>();
    let cursor = 0, curChunk = 0;
    results.push("");
    while(cursor < content.length) {
      const curChar = content[cursor];
      if (delimiters.includes(curChar)) {
        // check chunk length
        while (curChunk < results.length && results[curChunk].length >= length) {
          curChunk ++;
        }
        for (let i = curChunk; i < results.length; i++) {
          results[i] += curChar;
        }
        if (results[results.length - 1].length >= length - overlap) {
          results.push("");
        }
      } else {
        // append
        for (let i = curChunk; i < results.length; i++) {
          results[i] += curChar;
        }
      }
      cursor ++;
    }
    while (curChunk < results.length - 1) {
      results.pop();
    }
    return results;
  }
  
  

• embeddings.ts: Teams AI ライブラリ OpenAIEmbeddings を使用して埋め込みを作成します。

  import { OpenAIEmbeddings } from "@microsoft/teams-ai";
  
  const embeddingClient = new OpenAIEmbeddings({
    azureApiKey: "<your-aoai-key>",
    azureEndpoint: "<your-aoai-endpoint>",
    azureDeployment: "<your-embedding-deployment, e.g., text-embedding-ada-002>"
  });
  
  export async function createEmbeddings(content: string): Promise<number[]> {
    const response = await embeddingClient.createEmbeddings(content);
    return response.output[0];
  }
  

• searchIndex.ts: Azure AI Search インデックスを作成します。

  import { SearchIndexClient, AzureKeyCredential, SearchIndex } from "@azure/search-documents";
  
  const endpoint = "<your-search-endpoint>";
  const apiKey = "<your-search-key>";
  const indexName = "<your-index-name>";
  
  const indexDef: SearchIndex = {
    name: indexName,
    fields: [
      {
        type: "Edm.String",
        name: "id",
        key: true,
      },
      {
        type: "Edm.String",
        name: "content",
        searchable: true,
      },
      {
        type: "Edm.String",
        name: "filepath",
        searchable: true,
        filterable: true,
      },
      {
        type: "Collection(Edm.Single)",
        name: "contentVector",
        searchable: true,
        vectorSearchDimensions: 1536,
        vectorSearchProfileName: "default"
      }
    ],
    vectorSearch: {
      algorithms: [{
        name: "default",
        kind: "hnsw"
      }],
      profiles: [{
        name: "default",
        algorithmConfigurationName: "default"
      }]
    },
    semanticSearch: {
      defaultConfigurationName: "default",
      configurations: [{
        name: "default",
        prioritizedFields: {
          contentFields: [{
            name: "content"
          }]
        }
      }]
    }
  };
  
  export async function createNewIndex(): Promise<void> {
    const client = new SearchIndexClient(endpoint, new AzureKeyCredential(apiKey));
    await client.createIndex(indexDef);
  }
  

• searchIndexer.ts: 作成した埋め込みとその他のフィールドを Azure AI Search Index にアップロードします。

  import { AzureKeyCredential, SearchClient } from "@azure/search-documents";
  
  export interface Doc {
    id: string,
    content: string,
    filepath: string,
    contentVector: number[]
  }
  
  const endpoint = "<your-search-endpoint>";
  const apiKey = "<your-search-key>";
  const indexName = "<your-index-name>";
  const searchClient: SearchClient<Doc> = new SearchClient<Doc>(endpoint, indexName, new AzureKeyCredential(apiKey));
  
  export async function indexDoc(doc: Doc): Promise<boolean> {
    const response = await searchClient.mergeOrUploadDocuments([doc]);
    return response.results.every((result) => result.succeeded);
  }
  

• index.ts: 上記のコンポーネントを調整します。

  import { createEmbeddings } from "./embeddings";
  import { loadTextFile } from "./loader";
  import { createNewIndex } from "./searchIndex";
  import { indexDoc } from "./searchIndexer";
  import { split } from "./splitter";
  
  async function main() {
    // Only need to call once
    await createNewIndex();
  
    // local files as source input
    const files = [`${__dirname}/data/A.md`, `${__dirname}/data/A.md`];
    for (const file of files) {
      // load file
      const fullContent = loadTextFile(file);
  
      // split into chunks
      const contents = split(fullContent, 1000, 100);
      let partIndex = 0;
      for (const content of contents) {
        partIndex ++;
        // create embeddings
        const embeddings = await createEmbeddings(content);
  
        // upload to index
        await indexDoc({
          id: `${file.replace(/[^a-z0-9]/ig, "")}___${partIndex}`,
          content: content,
          filepath: file,
          contentVector: embeddings,
        });
      }
    }
  }
  
  main().then().finally();
  

• loader.py: ソース入力としてのプレーン テキスト。

  def load_text_file(path: str) -> str:
      with open(path, 'r', encoding='utf-8') as file:
          return file.read()
  

• splitter.py: テキストをチャンクに分割し、重複します。

  def split(content: str, length: int, overlap: int) -> list[str]:
      delimiters = [" ", "\t", "\r", "\n"]
      results = [""]
      cursor = 0
      cur_chunk = 0
      while cursor < len(content):
          cur_char = content[cursor]
          if cur_char in delimiters:
              while cur_chunk < len(results) and len(results[cur_chunk]) >= length:
                  cur_chunk += 1
              for i in range(cur_chunk, len(results)):
                  results[i] += cur_char
              if len(results[-1]) >= length - overlap:
                  results.append("")
          else:
              for i in range(cur_chunk, len(results)):
                  results[i] += cur_char
          cursor += 1
      while cur_chunk < len(results) - 1:
          results.pop()
      return results
  

• embeddings.py: Teams AI ライブラリ OpenAIEmbeddings を使用して埋め込みを作成します。

  async def create_embeddings(text: str, embeddings):
      result = await embeddings.create_embeddings(text)
  
      return result.output[0]
  

• search_index.py: Azure AI Search インデックスを作成します。

  async def create_index_if_not_exists(client: SearchIndexClient, name: str):
      doc_index = SearchIndex(
          name=name,
          fields = [
              SimpleField(name="docId", type=SearchFieldDataType.String, key=True),
              SimpleField(name="docTitle", type=SearchFieldDataType.String),
              SearchableField(name="description", type=SearchFieldDataType.String, searchable=True),
              SearchField(name="descriptionVector", type=SearchFieldDataType.Collection(SearchFieldDataType.Single), searchable=True, vector_search_dimensions=1536, vector_search_profile_name='my-vector-config'),
          ],
          scoring_profiles=[],
          cors_options=CorsOptions(allowed_origins=["*"]),
          vector_search = VectorSearch(
              profiles=[VectorSearchProfile(name="my-vector-config", algorithm_configuration_name="my-algorithms-config")],
              algorithms=[HnswAlgorithmConfiguration(name="my-algorithms-config")],
          )
      )
  
      client.create_or_update_index(doc_index)
  

• search_indexer.py: 作成した埋め込みとその他のフィールドを Azure AI Search Index にアップロードします。

  from embeddings import create_embeddings
  from search_index import create_index_if_not_exists
  from loader import load_text_file
  from split import split
  
  async def get_doc_data(embeddings):
      file_path=f'{os.getcwd()}/my_file_path_1'
      raw_description1 = split(load_text_file(file_path), 1000, 100)
      doc1 = {
          "docId": "1",
          "docTitle": "my_titile_1",
          "description": raw_description1,
          "descriptionVector": await create_embeddings(raw_description1, embeddings=embeddings),
      }
  
      file_path=f'{os.getcwd()}/my_file_path_2'
      raw_description2 = split(load_text_file(file_path), 1000, 100)
      doc2 = {
          "docId": "2",
          "docTitle": "my_titile_2",
          "description": raw_description2,
          "descriptionVector": await create_embeddings(raw_description2, embeddings=embeddings),
      }
  
      return [doc1, doc2]
  
  async def setup(search_api_key, search_api_endpoint):
      index = 'my_index_name'
      credentials = AzureKeyCredential(search_api_key)
      search_index_client = SearchIndexClient(search_api_endpoint, credentials)
      await create_index_if_not_exists(search_index_client, index)
  
      search_client = SearchClient(search_api_endpoint, index, credentials)
      embeddings=AzureOpenAIEmbeddings(AzureOpenAIEmbeddingsOptions(
            azure_api_key="<your-aoai-key>",
            azure_endpoint="<your-aoai-endpoint>",
            azure_deployment="<your-embedding-deployment, e.g., text-embedding-ada-002>"
      ))
      data = await get_doc_data(embeddings=embeddings)
      await search_client.merge_or_upload_documents(data)
  

• index.py: 上記のコンポーネントを調整します。

  from search_indexer import setup
  
  search_api_key = '<your-key>'
  search_api_endpoint = '<your-endpoint>'
  asyncio.run(setup(search_api_key, search_api_endpoint))
  

const { AzureKeyCredential, SearchClient } = require("@azure/search-documents");
const { DataSource, Memory, OpenAIEmbeddings, Tokenizer } = require("@microsoft/teams-ai");
const { TurnContext } = require("botbuilder");

// Define the interface for document
class Doc {
  constructor(id, content, filepath) {
    this.id = id;
    this.content = content; // searchable
    this.filepath = filepath;
  }
}

// Azure OpenAI configuration
const aoaiEndpoint = "<your-aoai-endpoint>";
const aoaiApiKey = "<your-aoai-key>";
const aoaiDeployment = "<your-embedding-deployment, e.g., text-embedding-ada-002>";

// Azure AI Search configuration
const searchEndpoint = "<your-search-endpoint>";
const searchApiKey = "<your-search-apikey>";
const searchIndexName = "<your-index-name>";

// Define MyDataSource class implementing DataSource interface
class MyDataSource extends DataSource {
  constructor() {
    super();
    this.name = "my-datasource";
    this.embeddingClient = new OpenAIEmbeddings({
      azureEndpoint: aoaiEndpoint,
      azureApiKey: aoaiApiKey,
      azureDeployment: aoaiDeployment
    });
    this.searchClient = new SearchClient(searchEndpoint, searchIndexName, new AzureKeyCredential(searchApiKey));
  }

  async renderData(context, memory, tokenizer, maxTokens) {
    // use user input as query
    const input = memory.getValue("temp.input");

    // generate embeddings
    const embeddings = (await this.embeddingClient.createEmbeddings(input)).output[0];

    // query Azure AI Search
    const response = await this.searchClient.search(input, {
      select: [ "id", "content", "filepath" ],
      searchFields: ["rawContent"],
      vectorSearchOptions: {
        queries: [{
          kind: "vector",
          fields: [ "contentVector" ],
          vector: embeddings,
          kNearestNeighborsCount: 3
        }]
      },
      queryType: "semantic",
      top: 3,
      semanticSearchOptions: {
        // your semantic configuration name
        configurationName: "default",
      }
    });

    // Add documents until you run out of tokens
    let length = 0, output = '';
    for await (const result of response.results) {
      // Start a new doc
      let doc = `${result.document.content}\n\n`;
      let docLength = tokenizer.encode(doc).length;
      const remainingTokens = maxTokens - (length + docLength);
      if (remainingTokens <= 0) {
          break;
      }

      // Append doc to output
      output += doc;
      length += docLength;
    }
    return { output, length, tooLong: length > maxTokens };
  }
}

async def get_embedding_vector(text: str):
    embeddings = AzureOpenAIEmbeddings(AzureOpenAIEmbeddingsOptions(
        azure_api_key='<your-aoai-key>',
        azure_endpoint='<your-aoai-endpoint>',
        azure_deployment='<your-aoai-embedding-deployment>'
    ))
    
    result = await embeddings.create_embeddings(text)
    if (result.status != 'success' or not result.output):
        raise Exception(f"Failed to generate embeddings for description: {text}")
    
    return result.output[0]

@dataclass
class Doc:
    docId: Optional[str] = None
    docTitle: Optional[str] = None
    description: Optional[str] = None
    descriptionVector: Optional[List[float]] = None

@dataclass
class MyDataSourceOptions:
    name: str
    indexName: str
    azureAISearchApiKey: str
    azureAISearchEndpoint: str

from azure.core.credentials import AzureKeyCredential
from azure.search.documents import SearchClient
import json

@dataclass
class Result:
    def __init__(self, output, length, too_long):
        self.output = output
        self.length = length
        self.too_long = too_long

class MyDataSource(DataSource):
    def __init__(self, options: MyDataSourceOptions):
        self.name = options.name
        self.options = options
        self.searchClient = SearchClient(
            options.azureAISearchEndpoint,
            options.indexName,
            AzureKeyCredential(options.azureAISearchApiKey)
        )
        
    def name(self):
        return self.name

    async def render_data(self, _context: TurnContext, memory: Memory, tokenizer: Tokenizer, maxTokens: int):
        query = memory.get('temp.input')
        embedding = await get_embedding_vector(query)
        vector_query = VectorizedQuery(vector=embedding, k_nearest_neighbors=2, fields="descriptionVector")

        if not query:
            return Result('', 0, False)

        selectedFields = [
            'docTitle',
            'description',
            'descriptionVector',
        ]

        searchResults = self.searchClient.search(
            search_text=query,
            select=selectedFields,
            vector_queries=[vector_query],
        )

        if not searchResults:
            return Result('', 0, False)

        usedTokens = 0
        doc = ''
        for result in searchResults:
            tokens = len(tokenizer.encode(json.dumps(result["description"])))

            if usedTokens + tokens > maxTokens:
                break

            doc += json.dumps(result["description"])
            usedTokens += tokens

        return Result(doc, usedTokens, usedTokens > maxTokens)