Wikipedia Semantic Search

This project demonstrates a full RAG (Retrieval-Augmented Generation) pipeline using Wikipedia data. The architecture consists of:

• Embedding Generation: Uses SentenceTransformers (specifically the BGE-M3 model) to create multilingual embeddings.
• Vector Storage & Search: Uses Upstash Vector to store and perform semantic searches across millions of articles.
• Session Management: Uses Upstash Redis to persist chat sessions.
• RAG Application: Built using the Upstash RAG Chat SDK.
• LLM Inference: Uses Meta-Llama-3-8B-Instruct via QStash LLM APIs to generate responses based on retrieved Wikipedia context.

To run this project locally, you need to configure Upstash databases and environment variables.

1. Provision Upstash Resources:

   • Create an Upstash Vector database. For multilingual support, select the BGE-M3 embedding model.
   • Create an Upstash Redis database to manage chat sessions.
   • Obtain QStash credentials to access hosted LLM models.

2. Configure Environment Variables: Create a .env file in the project root and populate it with your credentials:

3. Install and Run: Use pnpm to install dependencies and start the development server.

# .env file template
UPSTASH_VECTOR_REST_URL=
UPSTASH_VECTOR_REST_TOKEN=

UPSTASH_REDIS_REST_TOKEN=
UPSTASH_REDIS_REST_URL=

QSTASH_TOKEN=

pnpm install
pnpm dev

The project uses WikiMetadata to store descriptive information about a Wikipedia article and Wiki to represent the search result object containing the vector similarity score and the actual content.

• WikiMetadata contains the id, url, and title of the article.
• Wiki contains the id (number or string), the similarity score, optional metadata of type WikiMetadata, and optional data (the article content as a string).

export type WikiMetadata = {
  id: string;
  url: string;
  title: string;
};

export type Wiki = {
  id: number | string;
  score: number;
  metadata?: WikiMetadata | undefined;
  data?: string;
};

Perform a semantic search against the Upstash Vector index. This function automatically enhances the user's query by using OpenAI to extract relevant keywords, then queries the index using those keywords to improve retrieval accuracy. It uses the user's locale as a namespace for the query.

Returns: An object containing:

• code: ResultCode.Success if successful, or error codes like ResultCode.MinLengthError or ResultCode.UnknownError.
• data: The retrieved WikiMetadata results.
• ms: The time taken for the query in milliseconds.

import { serverQueryIndex } from '@/lib/actions';

const result = await serverQueryIndex("Who was Gandhi?");

if (result.code === 'Success') {
  console.log("Results:", result.data);
  console.log("Query time:", result.ms, "ms");
} else {
  console.error("Query failed with code:", result.code);
}

The service provides utilities to manage the user's preferred language (locale) using Next.js cookies. It supports a specific set of language codes and defaults to English if no locale is set.

Supported Locales: de, en, es, fa, fr, it, ja, pt, ru, tr, zh

Key Functions:

• getUserLocale(): An asynchronous function that retrieves the current locale from the NEXT_LOCALE cookie. If the cookie is missing, it returns the default locale (en).
• setUserLocale(locale: Locale): An asynchronous function that sets the NEXT_LOCALE cookie to the provided locale value.

import { getUserLocale, setUserLocale } from './service';

// Get the current user locale
const locale = await getUserLocale();

// Set a new locale for the user
await setUserLocale('fr');

Deletes all messages associated with the current sessionId found in the browser cookies.

import { serverClearMessages } from '@/lib/actions';

await serverClearMessages();

Retrieves metadata and information about the current Upstash Vector index state.

import { serverGetInfo } from '@/lib/actions';

const info = await serverGetInfo();
if (info) {
  console.log("Index Info:", info);
}

The project provides pre-initialized instances of the Upstash Vector Index and Upstash Redis clients.

• index: An instance of Index<WikiMetadata> used for semantic search operations. It requires the environment to be configured with Upstash Vector credentials.
• redis: An instance of the Upstash Redis client initialized via Redis.fromEnv(). This expects UPSTASH_REDIS_REST_URL and UPSTASH_REDIS_REST_TOKEN environment variables to be set.

import { index, redis } from './lib/dbs';

// Use the vector index for semantic search
// const results = await index.query(...);

// Use the redis client for key-value operations
// await redis.set('key', 'value');

The RAGChat class is initialized with a configuration object. Based on the implementation, the following keys are used:

• model: An LLM provider instance (e.g., using openai()). Requires an apiKey.
• vector: The Upstash Vector index instance used for semantic search.
• redis: The Redis instance used for managing chat history/state.
• debug: A boolean flag to enable/disable debug logging.
• promptFn: A function ({ chatHistory, context, question }) => string used to construct the final prompt sent to the model.