RAG Demo — You.com

Live demo →

A minimal, end-to-end Retrieval-Augmented Generation (RAG) demo built with Next.js. Upload a .txt file, ask a question, and get a grounded answer streamed back — powered by local embeddings and the You.com Express Agent.

---

How RAG Works

Large language models are powerful, but they only know what was in their training data. When you need answers grounded in your documents — not general knowledge — you need RAG.

RAG solves this in three steps:

① INDEXING (per request)
Your Document → [Chunk & Embed] → In-Memory Vector Store

② QUERYING
Your Question → [Embed] → Question Vector
                               ↓
                  Compare against Vector Store
                               ↓
                         Top Chunks
                    (most semantically similar passages)
                               ↓
③ Your Question + Top Chunks → [You.com Express Agent] → Streamed Answer

1. Index

The uploaded document is split into small, overlapping chunks of text. Each chunk is converted into a vector embedding — an array of numbers (e.g. [0.21, -0.87, 0.54, ...]) that represents the meaning of that text. Similar ideas produce similar numbers, which is what makes semantic search possible.

Unlike a traditional RAG system, this demo holds the index entirely in memory for the duration of the request — no database, no disk writes.

2. Retrieve

When a user asks a question, that question is run through the same embedding model to produce its own vector. The in-memory vector store compares it against all stored document vectors and returns the closest matches — the chunks most semantically related to the question.

This is why "What movies involve a kid and a father?" can surface content about Interstellar even if the word "father" never appears in the text.

3. Generate

The retrieved chunks are inserted into a prompt alongside the original question. The You.com Express Agent reads that context and streams a grounded answer back to the browser — without hallucinating facts it doesn't have.

---

Why You.com?

You.com's Express Agent is a fast, capable LLM endpoint designed for agentic and RAG workflows. In this pipeline it handles the generation step — it receives the retrieved context and question, then streams a precise, grounded answer.

Key advantages:

• Streaming by default — responses stream token-by-token, keeping latency low
• No hallucination pressure — explicitly prompted to answer only from provided context
• No mandatory web search — tools are opt-in, so the agent stays focused on your documents
• Simple API — one SDK call with the express agent

---

Project Structure

app/
  page.tsx            # UI: file upload, query input, streamed response
  api/
    query/
      route.ts        # API route: embed → retrieve → generate pipeline
public/
  classroom-favorite-movies.txt   # Sample document

---

How the Code Works

app/api/query/route.ts — Index, Retrieve & Generate

Accepts a POST request with { files, query, apiKey } and runs all three RAG steps server-side.

Embedding model: BAAI/bge-small-en-v1.5 runs entirely on-device via @llamaindex/huggingface — no data leaves your machine during the embedding step. The model is bundled with the deployment under /models so no download is needed at runtime.

In-memory index: Unlike a persistent RAG system, there's no disk storage. VectorStoreIndex.fromDocuments() is called without a storageContext, so the entire index lives in memory for the lifetime of the request. This keeps the architecture simple and stateless.

// No storageContext → stays entirely in memory, gone after the request
return VectorStoreIndex.fromDocuments(documents);

Chunking & retrieval: LlamaIndex's VectorStoreIndex handles splitting the document into chunks and cosine-similarity search. retrieve() returns the top-3 most relevant chunks, which are formatted into a numbered context block and injected into the You.com prompt.

Streaming: The You.com Express Agent streams its response token-by-token. The route forwards that stream directly to the browser as a text/plain ReadableStream:

const stream = await you.agentsRuns({ agent: "express", input, stream: true });

for await (const chunk of stream) {
  if (chunk.data.type === "response.output_text.delta") {
    controller.enqueue(encoder.encode(chunk.data.response.delta));
  }
}

app/page.tsx — UI

The browser handles everything else:

• API key input (passed directly in the request — no .env.local needed)
• File selection, including multiple files, via the File API (or loading the built-in example)
• Reading files as text and POSTing them to /api/query
• Reading the streamed response and appending tokens to the UI in real time

---

Getting Started

1. Install dependencies

npm install

2. Run the dev server

npm run dev

Open localhost:3000.

3. Try it

Enter your You.com API key (get one at you.com/platform), upload one or more .txt files (or click "Use our example"), and ask a question. The example file is a short story about a fifth-grade class sharing their favorite movies — try asking:

• "What is Elijah's favorite movie?"
• "Which students like animated films?"
• "Who relates most to their favorite movie character?"

---

Dependencies

Package	Purpose
next	App framework — UI and API routes
llamaindex	Document chunking, vector store, and similarity search
@llamaindex/huggingface	On-device embeddings via BAAI/bge-small-en-v1.5
@youdotcom-oss/sdk	You.com Express Agent for answer generation