google-gemini-file-search

Google Gemini File Search

Status: Production Ready | Last Verified: 2025-11-18

What Is File Search?

Google Gemini File Search is fully managed RAG (Retrieval-Augmented Generation):

• Upload documents → Automatic chunking + embeddings + vector search + citations
• No vector database setup required
• 100+ file formats supported (PDF, Word, Excel, code, Markdown, JSON, etc.)
• Built-in grounding with citation metadata
• Cost-effective: $0.15/1M tokens (one-time indexing), free storage + queries

Key difference from other RAG:

• Cloudflare Vectorize: You manage chunking/embeddings
• OpenAI Files API: Tied to Assistants API threads
• File Search: Fully managed, standalone RAG

Quick Start (5 Minutes)

1. Get API Key & Install

Get API key: https://aistudio.google.com/apikey (Free tier: 1 GB storage, 1,500 requests/day)

bun add @google/genai

Version: 0.21.0+ | Node.js: 18+

2. Basic Example

import { GoogleGenerativeAI } from '@google/genai';
import fs from 'fs';

const ai = new GoogleGenerativeAI(process.env.GOOGLE_AI_API_KEY);

// Create store
const fileStore = await ai.fileSearchStores.create({
  config: { displayName: 'my-knowledge-base' }
});

// Upload document
const operation = await ai.fileSearchStores.uploadToFileSearchStore({
  name: fileStore.name,
  file: fs.createReadStream('./manual.pdf'),
  config: {
    displayName: 'Installation Manual',
    chunkingConfig: {
      whiteSpaceConfig: {
        maxTokensPerChunk: 500,
        maxOverlapTokens: 50
      }
    }
  }
});

// Poll until done
while (!operation.done) {
  await new Promise(resolve => setTimeout(resolve, 1000));
  operation = await ai.operations.get({ name: operation.name });
}

// Query documents
const model = ai.getGenerativeModel({
  model: 'gemini-2.5-pro',  // Only 2.5 Pro/Flash supported
  tools: [{
    fileSearchTool: {
      fileSearchStores: [fileStore.name]
    }
  }]
});

const result = await model.generateContent('How do I install the product?');
console.log(result.response.text());

// Get citations
const grounding = result.response.candidates[0].groundingMetadata;
if (grounding) {
  console.log('Sources:', grounding.groundingChunks);
}

Load references/setup-guide.md for complete walkthrough with batch uploads, error handling, and production checklist.

Critical Rules

Always Do

1. Use delete + re-upload for updates (documents are immutable)
2. Calculate 3x storage (embeddings + metadata = ~3x file size)
3. Configure chunking (500 tokens for technical docs, 800 for prose)
4. Poll operations until done: true (with timeout)
5. Use force: true when deleting stores with documents
6. Use Gemini 2.5 models only (2.5-pro or 2.5-flash)
7. Keep metadata under 20 fields per document
8. Estimate indexing costs ($0.15/1M tokens one-time)

Never Do

1. Never try to update documents (no PATCH API exists)
2. Never assume storage = file size (it's 3x)
3. Never skip chunking config (defaults may not be optimal)
4. Never upload without polling (operation may still be processing)
5. Never delete without force if store has documents
6. Never use Gemini 1.5 models (File Search requires 2.5)
7. Never exceed 20 metadata fields (hard limit)
8. Never upload large files without cost estimate

Top 3 Errors Prevented

Error 1: Document Immutability

Problem: Trying to update existing document

Solution: Delete + re-upload pattern

// Find and delete old version
const docs = await ai.fileSearchStores.documents.list({
  parent: fileStore.name
});
const oldDoc = docs.documents.find(d => d.displayName === 'manual.pdf');
if (oldDoc) {
  await ai.fileSearchStores.documents.delete({
    name: oldDoc.name,
    force: true
  });
}

// Upload new version
await ai.fileSearchStores.uploadToFileSearchStore({
  name: fileStore.name,
  file: fs.createReadStream('manual-v2.pdf'),
  config: { displayName: 'manual.pdf' }
});

Error 2: Storage Quota Exceeded

Problem: Storage calculation wrong (3x multiplier)

Solution: Estimate before upload

const fileSize = fs.statSync('data.pdf').size;
const estimatedStorage = fileSize * 3;  // Embeddings + metadata

if (estimatedStorage > 1e9) {
  console.warn('⚠️ May exceed free tier 1 GB limit');
}

Error 3: Model Compatibility

Problem: Using wrong model version

Solution: Use Gemini 2.5 only

// ✅ CORRECT
const model = ai.getGenerativeModel({
  model: 'gemini-2.5-pro',  // or gemini-2.5-flash
  tools: [{ fileSearchTool: { fileSearchStores: [storeName] } }]
});

// ❌ WRONG
const model = ai.getGenerativeModel({
  model: 'gemini-1.5-pro',  // Not supported!
  tools: [{ fileSearchTool: { fileSearchStores: [storeName] } }]
});

Load references/error-catalog.md for all 8 errors with detailed solutions including chunking, operation polling, metadata limits, and force delete requirements.

When to Use File Search

Use File Search When:

• Want fully managed RAG (no vector DB)
• Cost predictability matters (one-time indexing)
• Need 100+ file format support
• Citations are important (built-in grounding)
• Simple deployment is priority
• Documents are relatively static

Use Alternatives When:

Cloudflare Vectorize - Global edge performance, custom embeddings, real-time R2 updates OpenAI Files API - Assistants API, conversational threads, very large collections (10,000+)

Common Patterns

Pattern 1: Customer Support Knowledge Base

// Upload support docs with metadata
await ai.fileSearchStores.uploadToFileSearchStore({
  name: fileStore.name,
  file: fs.createReadStream('troubleshooting.pdf'),
  config: {
    displayName: 'Troubleshooting Guide',
    customMetadata: {
      doc_type: 'support',
      category: 'troubleshooting',
      language: 'en'
    }
  }
});

Pattern 2: Batch Document Upload

const files = ['doc1.pdf', 'doc2.md', 'doc3.docx'];
const uploadPromises = files.map(file =>
  ai.fileSearchStores.uploadToFileSearchStore({
    name: fileStore.name,
    file: fs.createReadStream(file),
    config: { displayName: file }
  })
);
const operations = await Promise.all(uploadPromises);

// Poll all operations
for (const op of operations) {
  let operation = op;
  while (!operation.done) {
    await new Promise(resolve => setTimeout(resolve, 1000));
    operation = await ai.operations.get({ name: operation.name });
  }
  console.log('✅', operation.response.displayName);
}

Pattern 3: Document Update Flow

// 1. List existing documents
const docs = await ai.fileSearchStores.documents.list({
  parent: fileStore.name
});

// 2. Delete old version
const oldDoc = docs.documents.find(d => d.displayName === 'manual.pdf');
if (oldDoc) {
  await ai.fileSearchStores.documents.delete({
    name: oldDoc.name,
    force: true
  });
}

// 3. Upload new version
const operation = await ai.fileSearchStores.uploadToFileSearchStore({
  name: fileStore.name,
  file: fs.createReadStream('manual-v2.pdf'),
  config: {
    displayName: 'manual.pdf',
    customMetadata: {
      version: '2.0',
      updated_at: new Date().toISOString()
    }
  }
});

// 4. Poll until done
while (!operation.done) {
  await new Promise(resolve => setTimeout(resolve, 1000));
  operation = await ai.operations.get({ name: operation.name });
}

Load references/setup-guide.md for additional patterns including code documentation search and internal knowledge bases.

When to Load References

Load references/setup-guide.md when:

• First-time File Search setup
• Need step-by-step walkthrough with all configuration options
• Configuring batch upload strategies
• Production deployment checklist
• Complete API initialization patterns

Load references/error-catalog.md when:

• Encountering any of 8 common errors
• Need deta