Interviews RAG Actor

An Apify Actor that performs Retrieval-Augmented Generation (RAG) on meeting notes stored in Pinecone vector store. It retrieves relevant context and generates answers using GPT-4o.

This Actor runs only in Standby mode as an HTTP server for real-time API requests.

Features

• Vector Search: Retrieves relevant documents from Pinecone using similarity search
• GPT-4o Integration: Generates comprehensive answers using OpenAI's GPT-4o model
• Citation Support: Includes source citations with Notion URLs and dates
• Recency-Aware Ranking: Balances semantic similarity with document freshness
• Standby Mode: Runs as an HTTP server for real-time API requests
• Configurable: Adjustable retrieval parameters (k, threshold, recency)
• Per-Request Overrides: API keys, index name, LLM model, and all parameters can be overridden per request

Configuration

Default configuration is loaded from environment variables. All parameters can be overridden per-request.

Variable	Required	Description
OPENAI_API_KEY	✅	OpenAI API key for embeddings and LLM
PINECONE_API_KEY	✅	Pinecone API key
INDEX_NAME	✅	Pinecone index name containing embeddings
K	❌	Number of documents to retrieve (default: 20)
THRESHOLD	❌	Similarity threshold 0.0-1.0 (default: 0.3)
RECENCY_WEIGHT	❌	Balance between similarity and recency 0.0-1.0 (default: 0.2)
RECENCY_DECAY_DAYS	❌	Half-life for recency scoring in days (default: 180)
START_TEMPLATE	❌	Custom system prompt

Local Development

Create a .env file in the project root:

OPENAI_API_KEY=sk-your-openai-api-key
PINECONE_API_KEY=your-pinecone-api-key
INDEX_NAME=interviews

# Optional
K=20
THRESHOLD=0.3
RECENCY_WEIGHT=0.2

Apify Deployment

Configure environment variables in Actor Settings → Environment Variables on Apify Console.

API Endpoints

Method	Path	Description
GET	/ or /health	Health check and configuration status
POST	/ or /query	Submit a RAG query
GET	/query?question=...	Submit query via URL parameters

POST Request

Endpoint: POST /query or POST /

Headers:

Content-Type: application/json

Request Body:

{
    "question": "What did customers say about pricing?"
}

Per-request overrides:

All parameters below are optional. If not provided, values from environment variables are used.

Parameter	Type	Description
openai_api_key	string	Override OpenAI API key
pinecone_api_key	string	Override Pinecone API key
index_name	string	Override Pinecone index name
llm_model	string	LLM model (default: gpt-4o)
k	integer	Number of documents to retrieve (1-50)
threshold	number	Similarity threshold (0.0-1.0)
recency_weight	number	Recency vs similarity balance (0.0-1.0)
recency_decay_days	integer	Half-life for recency scoring
start_template	string	Custom system prompt

Example with overrides:

{
    "question": "What feedback did we get on the new feature?",
    "openai_api_key": "sk-your-key",
    "pinecone_api_key": "pc-your-key",
    "index_name": "custom-index",
    "llm_model": "gpt-4o-mini",
    "k": 5,
    "threshold": 0.7,
    "recency_weight": 0.3,
    "start_template": "Answer concisely based on the context."
}

Example with cURL:

curl -X POST https://your-actor.apify.actor/query \
  -H "Content-Type: application/json" \
  -d '{"question": "What are the main customer pain points?"}'

Response

{
    "answer": "Based on the meeting notes, customers mentioned several pain points...",
    "citations": [
        {
            "source_number": 1,
            "notion_url": "https://notion.so/...",
            "date": "2025-01-02"
        }
    ],
    "sources_used": 5
}

Error Response

{
    "error": "Missing 'question' field in request body",
    "error_type": "validation"
}

Local Development

# Create .env file with your credentials first

# Run locally in Standby mode
ACTOR_STANDBY_PORT=8080 apify run

# Deploy to Apify
apify login
apify push

How It Works

1. Loads configuration from environment variables at startup
2. Starts HTTP server listening for requests
3. For each query:
   • Connects to Pinecone and retrieves relevant documents
   • Applies recency-aware ranking to prioritize fresh content
   • Formats context with source metadata
   • Generates answer using GPT-4o with citation instructions
   • Extracts and returns citations with the response