RAG Cold Start Trap — $340 Embedding Bill in One Hour

import Parser from 'tree-sitter';
import { TypeScript } from 'tree-sitter-typescript';
import { OpenAIEmbeddings } from '@langchain/openai';
import { PgVectorStore } from '@langchain/community/vectorstores/pgvector';
import fs from 'fs/promises';
import crypto from 'crypto';
import { glob } from 'glob';

const IGNORED = ['**/node_modules/**', '**/.git/**', '**/.env*', '**/*.pem', '**/dist/**'];

const parser = new Parser();
parser.setLanguage(TypeScript.typescript);

const embeddings = new OpenAIEmbeddings({ model: 'text-embedding-3-large', batchSize: 100 });
const store = await PgVectorStore.initialize(embeddings, {
  postgresConnectionOptions: { connectionString: process.env.DATABASE_URL },
  tableName: 'code_embeddings'
});

export async function indexRepo() {
  const files = await glob('**/*.{ts,tsx}', { ignore: IGNORED });
  for (const path of files) {
    const source = await fs.readFile(path, 'utf-8');
    const tree = parser.parse(source);
    const nodes = tree.rootNode.descendantsOfType(['function_declaration', 'class_declaration', 'method_definition']);
    const imports = tree.rootNode.descendantsOfType('import_statement').map(n => source.slice(n.startIndex, n.endIndex)).join('\n');
    
    const docs = nodes.map(node => {
      const body = source.slice(node.startIndex, node.endIndex);
      const content = `${imports}\n\n${body}`;
      const hash = crypto.createHash('sha256').update(content).digest('hex');
      return { pageContent: content, metadata: { source: path, symbol: node.childForFieldName('name')?.text, hash } };
    });
    
    await store.addDocuments(docs, { ids: docs.map(d => d.metadata.hash) });
  }
}

import { streamText } from 'ai';
import { openai } from '@ai-sdk/openai';
import { PgVectorStore } from '@langchain/community/vectorstores/pgvector';
import { OpenAIEmbeddings } from '@langchain/openai';
import { Ratelimit } from '@upstash/ratelimit';
import { Redis } from '@upstash/redis';
import { encoding_for_model } from 'tiktoken';

const ratelimit = new Ratelimit({ redis: Redis.fromEnv(), limiter: Ratelimit.slidingWindow(100, '1 d') });
const embeddings = new OpenAIEmbeddings({ model: 'text-embedding-3-large' });
const store = await PgVectorStore.initialize(embeddings, { postgresConnectionOptions: { connectionString: process.env.DATABASE_URL! }, tableName: 'code_embeddings' });
const enc = encoding_for_model('gpt-4o');

function capContext(chunks: string[], maxTokens = 20000) {
  let total = 0; const kept = [];
  for (const c of chunks) { const t = enc.encode(c).length; if (total + t > maxTokens) break; kept.push(c); total += t; }
  return kept;
}

export async function POST(req: Request) {
  const { messages } = await req.json();
  const ip = req.headers.get('x-forwarded-for') ?? 'anon';
  const { success } = await ratelimit.limit(ip);
  if (!success) return new Response('Rate limited', { status: 429 });

  const query = messages.at(-1).content;
  const results = await store.similaritySearchWithScore(query, 12);
  const filtered = results.filter(([, score]) => score > 0.78);
  
  if (filtered.length === 0) return new Response("I don't see this in the codebase.");

  const context = capContext(filtered.map(([d]) => `[${d.metadata.source}]\n${d.pageContent}`)).join('\n\n---\n\n');

  const result = streamText({
    model: openai('gpt-4o'),
    system: `You are a coding assistant. Answer using ONLY this code. Cite files.\n\n${context}`,
    messages,
    temperature: 0.1,
  });

  return result.toTextStreamResponse();
}

import Parser from 'tree-sitter';
import { TypeScript } from 'tree-sitter-typescript';

const parser = new Parser();
parser.setLanguage(TypeScript.typescript);

export function chunkCode(source: string, filePath: string) {
  const tree = parser.parse(source);
  const imports = tree.rootNode.descendantsOfType('import_statement').map(n => source.slice(n.startIndex, n.endIndex)).join('\n');
  const nodes = tree.rootNode.descendantsOfType(['function_declaration', 'class_declaration', 'method_definition', 'arrow_function']);
  return nodes.map(node => ({
    content: `${imports}\n\n${source.slice(node.startIndex, node.endIndex)}`,
    metadata: { source: filePath, symbol: node.childForFieldName('name')?.text ?? 'anonymous' }
  })).filter(c => c.content.length > 50);
}

-- Neon pgvector hybrid search
SELECT content, metadata,
  1 - (embedding <=> $1) as vector_score,
  ts_rank(to_tsvector('english', content), plainto_tsquery($2)) as keyword_score
FROM code_embeddings
WHERE 1 - (embedding <=> $1) > 0.78
ORDER BY (0.7 * vector_score + 0.3 * keyword_score) DESC
LIMIT 12;