Student todo exercises

app/agents/linkedin.ts

@@ -1,26 +1,41 @@
-import { AgentRequest, AgentResponse } from './types';
-import { openai } from '@ai-sdk/openai';
-import { streamText } from 'ai';
+import { AgentRequest, AgentResponse } from "./types";
+import { openai } from "@ai-sdk/openai";
+import { streamText } from "ai";
 
 export async function linkedInAgent(
-	request: AgentRequest
+  request: AgentRequest,
 ): Promise<AgentResponse> {
-	// TODO: Step 1 - Get the fine-tuned model ID
-	// Access process.env.OPENAI_FINETUNED_MODEL
-	// If not configured, you might want to throw an error or use a fallback
+  // TODO: Step 1 - Get the fine-tuned model ID
+  // Access process.env.OPENAI_FINETUNED_MODEL
+  // If not configured, you might want to throw an error or use a fallback
+  const modelId = process.env.OPENAI_FINETUNED_MODEL;
+  if (!modelId) throw new Error("couldn't get model id");
 
-	// TODO: Step 2 - Build the system prompt
-	// Include instructions for the LinkedIn agent
-	// Add context about the original and refined queries:
-	//   - request.originalQuery - what the user originally asked
-	//   - request.query - the refined/improved version
-	// Tell the model to create engaging LinkedIn posts
+  // TODO: Step 2 - Build the system prompt
+  // Include instructions for the LinkedIn agent
+  // Add context about the original and refined queries:
+  //   - request.originalQuery - what the user originally asked
+  //   - request.query - the refined/improved version
+  // Tell the model to create engaging LinkedIn posts
+  const systemPrompt = `You are a LinkedIn agent. Analyze the users refined query and create a unique Linkedin post.
 
-	// TODO: Step 3 - Stream the response
-	// Use streamText() from the 'ai' package
-	// Pass the model using openai()
-	// Include system prompt and messages from request.messages
-	// Return the stream
+Your task:
+1. analyze the query
+2. identify the topic for the post
+3. create a unique post for the topic 
+4. return the new post.`;
+  // TODO: Step 3 - Stream the response
+  // Use streamText() from the 'ai' package
+  // Pass the model using openai()
+  // Include system prompt and messages from request.messages
+  // Return the stream
 
-	throw new Error('LinkedIn agent not implemented yet!');
+  return streamText({
+    model: openai(modelId),
+    system: systemPrompt,
+    prompt: `
+    Original User Request: ${request.originalQuery}
+    Refined Query: ${request.query}
+  `,
+  });
 }

app/agents/rag.ts

@@ -1,54 +1,100 @@
-import { AgentRequest, AgentResponse } from './types';
-import { pineconeClient } from '@/app/libs/pinecone';
-import { openaiClient } from '@/app/libs/openai/openai';
-import { openai } from '@ai-sdk/openai';
-import { streamText } from 'ai';
+import { AgentRequest, AgentResponse } from "./types";
+import { pineconeClient } from "@/app/libs/pinecone";
+import { openaiClient } from "@/app/libs/openai/openai";
+import { openai } from "@ai-sdk/openai";
+import { streamText } from "ai";
 
 export async function ragAgent(request: AgentRequest): Promise<AgentResponse> {
-	// TODO: Step 1 - Generate embedding for the refined query
-	// Use openaiClient.embeddings.create()
-	// Model: 'text-embedding-3-small'
-	// Dimensions: 512
-	// Input: request.query
-	// Extract the embedding from response.data[0].embedding
-
-	// TODO: Step 2 - Query Pinecone for similar documents
-	// Get the index: pineconeClient.Index(process.env.PINECONE_INDEX!)
-	// Query parameters:
-	//   - vector: the embedding from step 1
-	//   - topK: 10 (to over-fetch for reranking)
-	//   - includeMetadata: true
-
-	// TODO: Step 3 - Extract text from results
-	// Map over queryResponse.matches
-	// Get metadata?.text (or metadata?.content as fallback)
-	// Filter out any null/undefined values
-
-	// TODO: Step 4 - Rerank with Pinecone inference API
-	// Use pineconeClient.inference.rerank()
-	// Model: 'bge-reranker-v2-m3'
-	// Pass the query and documents array
-	// This gives you better quality results
-
-	// TODO: Step 5 - Build context from reranked results
-	// Map over reranked.data
-	// Extract result.document?.text from each
-	// Join with '\n\n' separator
-
-	// TODO: Step 6 - Create system prompt
-	// Include:
-	//   - Instructions to answer based on context
-	//   - Original query (request.originalQuery)
-	//   - Refined query (request.query)
-	//   - The retrieved context
-	//   - Instruction to say if context is insufficient
-
-	// TODO: Step 7 - Stream the response
-	// Use streamText()
-	// Model: openai('gpt-4o')
-	// System: your system prompt
-	// Messages: request.messages
-	// Return the stream
-
-	throw new Error('RAG agent not implemented yet!');
+  // TODO: Step 1 - Generate embedding for the refined query
+  // Use openaiClient.embeddings.create()
+  // Model: 'text-embedding-3-small'
+  // Dimensions: 512
+  // Input: request.query
+  // Extract the embedding from response.data[0].embedding
+  const response = await openaiClient.embeddings.create({
+    model: "text-embedding-3-small",
+    dimensions: 512,
+    input: request.query,
+  });
+  const embedding = response.data[0].embedding;
+  // TODO: Step 2 - Query Pinecone for similar documents
+  // Get the index: pineconeClient.Index(process.env.PINECONE_INDEX!)
+  // Query parameters:
+  //   - vector: the embedding from step 1
+  //   - topK: 10 (to over-fetch for reranking)
+  //   - includeMetadata: true
+  const index = pineconeClient.Index(process.env.PINECONE_INDEX!);
+  const result = await index.query({
+    vector: embedding,
+    topK: 10,
+    includeMetadata: true,
+  });
+
+  // TODO: Step 3 - Extract text from results
+  // Map over queryResponse.matches
+  // Get metadata?.text (or metadata?.content as fallback)
+  // Filter out any null/undefined values
+  const text = result.matches
+    .map((res) => res.metadata?.text)
+    .filter((text) => typeof text === "string");
+
+  // TODO: Step 4 - Rerank with Pinecone inference API
+  // Use pineconeClient.inference.rerank()
+  // Model: 'bge-reranker-v2-m3'
+  // Pass the query and documents array
+  // This gives you better quality results
+  const reranked = await pineconeClient.inference.rerank(
+    "bge-reranker-v2-m3",
+    request.query,
+    text,
+    { topN: 5, returnDocuments: true },
+  );
+
+  // TODO: Step 5 - Build context from reranked results
+  // Map over reranked.data
+  // Extract result.document?.text from each
+  // Join with '\n\n' separator
+  const retrievedContext = reranked.data
+    .map((result) => result.document?.text)
+    .filter(Boolean)
+    .join("\n\n");
+
+  // TODO: Step 6 - Create system prompt
+  // Include:
+  //   - Instructions to answer based on context
+  //   - Original query (request.originalQuery)
+  //   - Refined query (request.query)
+  //   - The retrieved context
+  //   - Instruction to say if context is insufficient
+  const systemPrompt = `You are a helpful assistant answering questions based on the provided context in the user's query.
+
+use the context in the query to answer the user's question.
+
+original query: ${request.originalQuery}
+refined query: ${request.query}
+retrieved context: ${retrievedContext}
+
+if the context doesnt contain enough information, say so clearly or do your best to answer the question`;
+
+  // TODO: Step 7 - Stream the response
+  // Use streamText()
+  // Model: openai('gpt-4o')
+  // System: your system prompt
+  // Messages: request.messages
+  // Return the stream
+  console.log(
+    "Pinecone scores:",
+    result.matches.map((m) => ({ score: m.score, text: m.metadata?.text })),
+  );
+  console.log(
+    "Re-ranked scores:",
+    reranked.data.map((r) => ({ score: r.score, text: r.document })),
+  );
+  console.log("Context length:", retrievedContext.length);
+
+  return streamText({
+    model: openai("gpt-4o"),
+    system: systemPrompt,
+    prompt: `Context: ${retrievedContext}\n\nQuery: ${request.query}`,
+  });
 }

app/api/rag-test/route.ts

@@ -1,24 +1,36 @@
-import { searchDocuments } from '@/app/libs/pinecone';
-import { NextRequest, NextResponse } from 'next/server';
+import { searchDocuments } from "@/app/libs/pinecone";
+import { NextRequest, NextResponse } from "next/server";
 
 export async function POST(request: NextRequest) {
-	const body = await request.json();
-	const { query, topK } = body;
+  try {
+    const body = await request.json();
+    const { query, topK = 5 } = body;
 
-	const results = await searchDocuments(query, topK);
+    if (!query || typeof query !== "string") {
+      console.error("Query is required for RAG search");
+      return NextResponse.json({ error: "Query is required" }, { status: 400 });
+    }
+    const results = await searchDocuments(query, topK);
 
-	const formattedResults = results.map((doc) => ({
-		id: doc.id,
-		score: doc.score,
-		content: doc.metadata?.text || '',
-		source: doc.metadata?.source || 'unknown',
-		chunkIndex: doc.metadata?.chunkIndex,
-		totalChunks: doc.metadata?.totalChunks,
-	}));
+    const formattedResults = results.map((doc) => ({
+      id: doc.id,
+      score: doc.score,
+      content: doc.metadata?.text || "",
+      source: doc.metadata?.source || "unknown",
+      chunkIndex: doc.metadata?.chunkIndex,
+      totalChunks: doc.metadata?.totalChunks,
+    }));
 
-	return NextResponse.json({
-		query,
-		resultsCount: formattedResults.length,
-		results: formattedResults,
-	});
+    return NextResponse.json({
+      query,
+      resultsCount: formattedResults.length,
+      results: formattedResults,
+    });
+  } catch (error) {
+    console.error("❌ Error in RAG test route:", error);
+    return NextResponse.json(
+      { error: "Failed to perform RAG search" },
+      { status: 500 },
+    );
+  }
 }