How to Build a VAPI Knowledge Base Using Gemini File Search Store

What is the Best Way to Connect VAPI to a Gemini RAG Pipeline?

A Step-by-Step Guide for Gemini File Search and Make.com Integration

Who is this Article for? | Article Use Case

This tutorial is for Voice AI developers and automation specialists in South Africa and beyond who use VAPI and Make.com. If you need your voice agent to retrieve specific facts from your own documents during a live call, this workflow is for you.

In this Article

• How Do You Reduce RAG Latency and Cost in VAPI?
• How Does Gemini File Search Enable RAG for Voice AI?
• What Are the Requirements for Gemini and Make.com Integration?
• Where Do I Get a Google Gemini API Key for My Agent?
• How Do You Build the Make.com Workflow for VAPI Retrieval?
• How to Upload Knowledge Base Files to the Gemini API?
• How Do You Create a File Search Store Using HTTP Requests?
• How Do You Link Uploaded Files to a Gemini Search Store?
• Why Are Metadata and Chunking Critical for Voice AI Accuracy?
• How Can You Apply These Solutions?

How Do You Reduce RAG Latency and Cost in VAPI?

The reason traditional RAG creates a bottleneck for voice agents is simple: it requires too many moving parts. Managing separate vector databases and embedding models adds latency that can make a voice call feel unnatural. By using the Gemini File Search Tool, you move to a "Managed RAG" model. This setup keeps the retrieval process within the Google infrastructure, which helps reduce response delays. It also cuts token costs by up to 70% because the model only retrieves specific sections of your data instead of processing an entire document for every query.

How Does Gemini File Search Enable RAG for Voice AI?

The Gemini API uses Retrieval Augmented Generation (RAG) to provide context to the model. This allows the model to provide more accurate answers based on your specific files. Understanding these three parts is key:

File Search

The core ability of Gemini to look through data to find relevant information based on a prompt.

File Search Tool

The tool that imports, chunks, and indexes your data for fast retrieval.

File Search Store

The container where your files are organized. You must create a store before you can search your files.

What Are the Requirements for Gemini and Make.com Integration?

You need a combination of API access and automation tools to bridge the gap between your voice agent and your data. Ensure you have the following ready:

• Google Gemini API Key: Required for all HTTP requests to Google's generative language services.
• Make.com Account: Used to orchestrate the flow between GitLab, Google Drive, and Gemini.
• VAPI Account: Your voice agent platform where the "Tool Call" will be configured.
• n8n Account: (Optional) An alternative to Make.com for those who prefer self-hosted automation.

Where Do I Get a Google Gemini API Key for My Agent?

You can generate your key through the Google AI Studio. The process is simple:

1. Go to the Google AI Studio page.
2. Click Get API Key in the bottom-left sidebar.
3. Select Create API Key at the top-right of the screen.
4. Copy the key and save it in a secure location for use in your Make.com scenario.

Image

How Do You Build the Make.com Workflow for VAPI Retrieval?

Building a robust RAG workflow requires a two-phase setup: one-off administrative tasks and a live operational scenario. The administrative phase involves:

• uploading your knowledge base (once-off);
• creating the search store (once-off), and;
• associating the knowledge base with the store (once-off).

The operational phase handles the live "Tool Call" from VAPI, retrieves the answer, and formats the JSON response.

How to Upload Knowledge Base Files to the Gemini API?

Uploading your file uploads from a repository on GitLab/GitHub ensures your voice agent always uses the most current data. While you can upload files manually, we would look to build a scenario that automates the maintenance process.

Scenario to Upload via GitLab (one-off)

1. Create GitLab Connection: Use a Personal Access Token (PAT) with API access.
2. Decode Content: Convert the GitLab file return from Base64 to a Markdown String.
3. Store Temporarily: Upload and retrieve from Google Drive.
4. Execute HTTP POST: Send the file to Gemini using the following configuration:
   1. Authentication type:API Key
   2. URL: https://generativelanguage.googleapis.com/upload/v1beta/files
   3. Method: POST
   4. Query Parameter: key = Your Gemini API Key
   5. Body Content Type: multipart/form-data
   6. Body content:
      1. Name: file
      2. Field type: File
      3. File: Map to the download file from Google Drive.

Typical Response:
[
    {
        "data": {
            "file": {
                "name": "files/SOME_ID",
                "mimeType": "text/markdown",
                "sizeBytes": "6180",
                "createTime": "2025-12-04T10:03:15.321360Z",
                "updateTime": "2025-12-04T10:03:15.321360Z",
                "expirationTime": "2025-12-06T10:03:14.459300427Z",
                "sha256Hash": "THE_HASH",
                "uri": "https://generativelanguage.googleapis.com/v1beta/files/SOME_ID",
                "state": "ACTIVE",
                "source": "UPLOADED"
            }
        },
        "headers": {
          .......
        },
        "statusCode": 200
    }
]

How-To Image Sequence for Upload File to Gemini

The sequence below gives a visual guide to the Scenario to Upload Via GitLab.

Flow Diagram

Image

The Base64 conversion:
{{toString(toBinary(12.content; "base64"))}}

The HTTP POST Configuration

Image

How Do You Create a File Search Store Using HTTP Requests?

Once your file is uploaded, you need a store to hold it. Use the HTTP Request module with these settings (once-off):

• Authentication type:API Key
• URL: https://generativelanguage.googleapis.com/v1beta/fileSearchStores
• Method: POST
• Header: Content-Type = application/json
• Query Parameter: key = Your Gemini API Key
• Body content type: application/json
• Body input method: JSON string
• Body Content: { "displayName": "name_your_store" }

Typical response:
[
   {
       "data": {
           "name": "fileSearchStores/STORE_NAME-STORE_ID",
           "displayName": "YOUR_DISPLAY_NAME_ENTERED",
           "createTime": "2025-12-04T10:10:53.022151Z",
           "updateTime": "2025-12-04T10:10:53.022151Z"
       },
       "headers": {
           ....
       },
       "statusCode": 200
   }
]

Image

How Do You Link Uploaded Files to a Gemini Search Store?

The final setup step is linking your file to the store. Send a POST request using the HTTP module:

• Authentication type:API Key
• URL: https://generativelanguage.googleapis.com/v1beta/fileSearchStores/YOUR-…
• Method: POST
• Header: Content-Type = application/json
• Query Parameter: key = Your Gemini API Key
• Body content type: application/json
• Body input method: JSON string

• Body Content:
  {
    "fileName": "{{FILE-FROM-UPLOAD-FILE}}",
    "customMetadata": [
      { "key": "topic", "stringValue": "general_questions" },
      { "key": "voice_style", "stringValue": "fast_track" },
      { "key": "source_version", "stringValue": "2025-11-15" }
    ],
    "chunkingConfig": {
      "whiteSpaceConfig": {
        "maxTokensPerChunk": 400,
        "maxOverlapTokens": 70
      }
    }
  }

Typical response:
[
   {
       "data": {
           "name": "fileSearchStores/STORE-ID/operations/ID-ID",
           "response": {
               "@type": "type.googleapis.com/google.ai.generativelanguage.v1main.ImportFileResponse"
           }
       },
       "headers": {
           ....
       },
       "statusCode": 200
   }
]

Image

Why Are Metadata and Chunking Critical for Voice AI Accuracy?

The reason is simple: it helps the AI find the needle in the haystack. Without these, the AI might retrieve too much or too little information.

What is customMetadata?

Metadata acts as a label. By tagging files with topics like "pricing" or "support," you help the AI prioritize the right documents for specific questions.

What is chunkingConfig?

Chunking breaks large files into small pieces. For voice agents, you want short chunks. We recommend a maxTokensPerChunk of 400. This ensures the answer is concise enough for a phone conversation.

KB Tool Set Up Coming Soon

I am taking a few days to visit family, the final Make.com scenario will be done between 27 - 31 December.

How Can You Apply These Solutions?

By following this structured approach, you turn a standard voice bot into an authoritative assistant. Use the File Search Store to organize your brand's Knowledge Nodes. This ensures that when a customer asks a question in South Africa or anywhere else, your VAPI agent responds with the exact information you've vetted. Start by setting up your API key and creating your first store today.

TL;DR

To give your VAPI agent a knowledge base, use Gemini's File Search Store. Upload files via HTTP in Make.com, create a store, and link the files using the importFile method. Use customMetadata for better organization and chunkingConfig to ensure the AI retrieves short, relevant answers perfect for voice conversations.

Links and downloads:

TODO