Vectara
Vectara 是一个用于构建 GenAI 应用程序的平台。它提供了一个易于使用的 API,用于文档索引和查询,该 API 由 Vectara 管理,并针对性能和准确性进行了优化。
您可以将 Vectara 用作 LangChain.js 的向量数据库。
👉 包含嵌入
Vectara 在后台使用自己的嵌入,因此您无需自己提供任何嵌入或调用其他服务来获取嵌入。
这也意味着,如果您提供自己的嵌入,它们将成为一个空操作。
const store = await VectaraStore.fromTexts(
["hello world", "hi there"],
[{ foo: "bar" }, { foo: "baz" }],
// This won't have an effect. Provide a FakeEmbeddings instance instead for clarity.
new OpenAIEmbeddings(),
args
);
设置
您需要
- 创建一个 免费的 Vectara 帐户。
- 创建一个 语料库 来存储您的数据
- 创建一个 API 密钥,具有 QueryService 和 IndexService 访问权限,以便您可以访问此语料库
配置您的 .env
文件或提供参数以将 LangChain 连接到您的 Vectara 语料库
VECTARA_CUSTOMER_ID=your_customer_id
VECTARA_CORPUS_ID=your_corpus_id
VECTARA_API_KEY=your-vectara-api-key
请注意,您可以提供以逗号分隔的多个语料库 ID 来同时查询多个语料库。例如:VECTARA_CORPUS_ID=3,8,9,43
。要索引多个语料库,您需要为每个语料库创建一个单独的 VectaraStore 实例。
用法
import { VectaraStore } from "@langchain/community/vectorstores/vectara";
import { VectaraSummaryRetriever } from "@langchain/community/retrievers/vectara_summary";
import { Document } from "@langchain/core/documents";
// Create the Vectara store.
const store = new VectaraStore({
customerId: Number(process.env.VECTARA_CUSTOMER_ID),
corpusId: Number(process.env.VECTARA_CORPUS_ID),
apiKey: String(process.env.VECTARA_API_KEY),
verbose: true,
});
// Add two documents with some metadata.
const doc_ids = await store.addDocuments([
new Document({
pageContent: "Do I dare to eat a peach?",
metadata: {
foo: "baz",
},
}),
new Document({
pageContent: "In the room the women come and go talking of Michelangelo",
metadata: {
foo: "bar",
},
}),
]);
// Perform a similarity search.
const resultsWithScore = await store.similaritySearchWithScore(
"What were the women talking about?",
1,
{
lambda: 0.025,
}
);
// Print the results.
console.log(JSON.stringify(resultsWithScore, null, 2));
/*
[
[
{
"pageContent": "In the room the women come and go talking of Michelangelo",
"metadata": {
"lang": "eng",
"offset": "0",
"len": "57",
"foo": "bar"
}
},
0.4678752
]
]
*/
const retriever = new VectaraSummaryRetriever({ vectara: store, topK: 3 });
const documents = await retriever.invoke("What were the women talking about?");
console.log(JSON.stringify(documents, null, 2));
/*
[
{
"pageContent": "<b>In the room the women come and go talking of Michelangelo</b>",
"metadata": {
"lang": "eng",
"offset": "0",
"len": "57",
"foo": "bar"
}
},
{
"pageContent": "<b>In the room the women come and go talking of Michelangelo</b>",
"metadata": {
"lang": "eng",
"offset": "0",
"len": "57",
"foo": "bar"
}
},
{
"pageContent": "<b>In the room the women come and go talking of Michelangelo</b>",
"metadata": {
"lang": "eng",
"offset": "0",
"len": "57",
"foo": "bar"
}
}
]
*/
// Delete the documents.
await store.deleteDocuments(doc_ids);
API 参考
- VectaraStore 来自
@langchain/community/vectorstores/vectara
- VectaraSummaryRetriever 来自
@langchain/community/retrievers/vectara_summary
- Document 来自
@langchain/core/documents
请注意,lambda
是与 Vectara 的混合搜索功能相关的参数,它提供了神经搜索和布尔/精确匹配之间的权衡,如 这里 所述。我们建议默认值为 0.025,同时为高级用户提供了一种方法,以便在需要时自定义此值。
APIs
Vectara 的 LangChain 向量数据库使用 Vectara 的核心 API