Nepali News Intelligence — Backend

A RAG-powered chatbot backend for querying Nepali financial news and NEPSE stock data. Built with FastAPI, PostgreSQL (pgvector), and Google Gemini, it serves an HTMX-driven UI with real-time chat, company insights, trending topics, and a watchlist.

This is the backend application. The data pipeline that scrapes, enriches, and populates the database lives in a separate repo: 👉 news-data-pipeline

---

Architecture

news-data-pipeline (separate repo)
        │
        ▼
  PostgreSQL (pgvector)        ◄──── This backend reads from here
        │
        ▼
  FastAPI Backend (this repo)
        │
        ▼
  Jinja2 + HTMX Frontend

The pipeline scrapes 14 Nepali news sources hourly, generates Gemini embeddings, and writes enriched articles into PostgreSQL. This backend queries that data to power the chatbot and insights features.

---

Features

• RAG Chat — Ask questions about Nepali financial news; answers are grounded in real articles via cosine similarity search over pgvector embeddings
• Trending Topics — Auto-generated hourly from high-signal articles
• Company Insights — Per-symbol NEPSE page with price chart, key stats, technical indicators, and AI-generated buy/hold/sell recommendations
• Watchlist — Track NEPSE companies and receive AI-generated impact summaries based on latest news
• Multi-tenant Org System — Organizations with admin/user roles, invite-code-based onboarding, and feature-level access control
• Admin Dashboard — Manage team members, assign features, and change roles
• Scheduled Scraper Trigger — Fires the pipeline's scraper queue via RabbitMQ every hour (skips midnight–5 AM Kathmandu time)

---

Tech Stack

Layer	Technology
Web framework	FastAPI
Frontend	Jinja2 templates + HTMX
Database	PostgreSQL with pgvector
AI / Embeddings	Google Gemini
Object storage	MinIO / Wasabi (S3-compatible)
Message queue	RabbitMQ
Package manager	Poetry
Runtime	Python 3.11

---

Prerequisites

• Python 3.11+
• Poetry
• PostgreSQL with the pgvector extension enabled
• A running RabbitMQ instance
• A MinIO or Wasabi bucket (used by the pipeline; credentials needed here for scraper triggers)
• Google Gemini API key

The database schema is created automatically on first startup via DbService.

---

Setup

1. Clone the repo

git clone https://github.com/Srmaraghu/news-backend.git
cd news-backend

2. Install dependencies

poetry install

3. Configure environment variables

cp .env.example .env
# Fill in all values in .env

4. Run the app

poetry run uvicorn src.api:app --host 0.0.0.0 --port 8000 --reload

The app will be available at http://localhost:8000.

---

Docker

docker build -t news-backend .
docker run --env-file .env -p 8000:8000 news-backend

Make sure your .env file is populated before running the container.

---

Environment Variables

Copy .env.example to .env and fill in your values.

Variable	Description
SECRET_KEY	Session signing key — use a long random string
DB_HOST, DB_PORT, DB_NAME, DB_USER, DB_PASSWORD	PostgreSQL connection
GEMINI_API_KEY	Google Gemini API key
GEMINI_EMBED_MODEL	Embedding model name (e.g. models/text-embedding-004)
GEMINI_GENERATE_MODEL	Generation model name (e.g. gemini-1.5-flash)
WASABI_ENDPOINT, WASABI_ACCESS_KEY, WASABI_SECRET_KEY	Object storage credentials
RABBITMQ_HOST, RABBITMQ_PORT, RABBITMQ_VHOST, RABBITMQ_USER, RABBITMQ_PASSWORD	RabbitMQ connection

---

Project Structure

src/
├── api.py                  # App entry point, router registration
├── config.py               # Middleware and template configuration
├── models.py               # Pydantic models
├── endpoints/
│   ├── auth.py             # Login, register, invite, team management
│   ├── conversation.py     # Chat / RAG query endpoint
│   ├── conversation_group.py
│   ├── company_insights.py # NEPSE company data + AI recommendations
│   ├── trending_topics.py
│   ├── watchlist.py
│   └── scraper_trigger.py  # Manual scraper trigger
├── services/
│   ├── embeddings/
│   │   ├── db_service.py       # All database operations
│   │   ├── gemini_client.py    # Gemini API wrapper
│   │   ├── gemini_response.py  # RAG answer generation
│   │   └── minio_service.py    # Object storage client
│   ├── prompts/                # YAML prompt templates
│   ├── schedule_service.py     # Hourly scraper scheduler
│   ├── trigger_scraper_service.py
│   ├── rabbitmq_service.py
│   ├── online_khabar_scraper.py
│   └── markdown_service.py
├── middlewares/
│   ├── helpers.py          # Auth guards (requires_role, requires_feature)
│   └── exception_handlers.py
├── templates/              # Jinja2 HTML templates
└── static/                 # CSS

---

User Roles & Access Control

Role	Capabilities
admin	Full access — manage team, assign features, generate invites
user	Access to features assigned by their org admin

Features (chat_box, company_insights, trending_topics) are toggled per-user by admins from the dashboard.

---

Related Repos

• Data Pipeline → Srmaraghu/news-data-pipeline — Scrapes 14 Nepali news sources, generates embeddings, and populates the PostgreSQL database that this backend reads from.

---

Blog Posts

• Building a Context-Aware Financial News Chatbot — Deep dive into how the RAG pipeline works: time-aware query parsing, cosine similarity retrieval over 768-dim Gemini embeddings, prompt assembly with conversation history, and the Gemini financial analyst persona.

---

Notes

• SSL certificate verification is disabled in rabbitmq_service.py for development. Enable proper cert validation before deploying to production.
• The scraper scheduler skips runs between midnight and 5 AM Kathmandu time (Asia/Kathmandu).