Intent Detection Agent

Intent Detection Agent is an open-source system for discovering B2B sales prospects from live web signals. It transforms natural language queries into ranked company lists by orchestrating web search, signal ingestion, classification, and scoring in a modular multi-agent pipeline.

Key Features

• Multi-agent pipeline: Structured workflow using LangGraph
• Knowledge graph integration: Neo4j-based company–signal relationships
• Semantic search: Vector retrieval with Qdrant
• Explainable scoring: Transparent fit-score computation
• API-based deployment: FastAPI backend for real-world integration

---

System Architecture

The system consists of four main stages:

1. Web Search – Collects real-time signals from the web using Perplexity API
2. Ingestion – Cleans, deduplicates, and stores signals (Neo4j + Qdrant)
3. Classification – Assigns signal type and sentiment using LLM + rules
4. Scoring – Aggregates signals into company-level fit scores

Pipeline flow:

web_search → ingest → classify → score

---

Impact

• Reduces manual prospect research from 2–3 hours → < 1 minute
• Estimated cost per query: ~$0.11
• Enables scalable B2B lead generation and market intelligence workflows

---

Tech Stack

• Backend: Python, FastAPI
• Orchestration: LangGraph
• Databases: Neo4j (graph), Qdrant (vector)
• APIs: OpenAI, Perplexity
• Frontend: Next.js, TypeScript

---

Setup Instructions

Prerequisites

• Python 3.11+

• Node.js 18+

• PostgreSQL-compatible database

• API keys for:

  • OpenAI
  • Perplexity

Installation

git clone https://github.com/1Ninad/Intent-Detection-Agent.git
cd Intent-Detection-Agent

Install backend dependencies:

pip install -r requirements.txt

Run backend:

uvicorn app.main:app --reload

Run frontend:

cd frontend
npm install
npm run dev

---

Repository Structure

backend/
frontend/
evaluation/

• backend/ – Core pipeline and API
• frontend/ – UI for query input and results
• evaluation/ – Sample outputs and evaluation artifacts

---

Evaluation

• Classification accuracy: 63.3%
• Sentiment accuracy: 96.7%
• Average cost: ~$0.11/query

---

Limitations

• Depends on external APIs (OpenAI, Perplexity)
• Performance may vary with web signal quality
• Latency primarily influenced by web search stage

---

Future Work

• Improve classification accuracy for minority signal types
• Reduce end-to-end latency
• Expand domain-specific signal taxonomies
• Improve reproducibility packaging

---

License

This project is licensed under the MIT License.

---