philipakintola01-sys/naija-tax-ai
Domaine:
digital infrastructureType de record:
software
Créateur:
phi
Hôte:
RAG-powered Nigerian tax AI agent — answers tax questions with citations from official 2025 tax reform documents (NTA, NTAA, NRSA, JRBA) via Telegram. Built with n8n, Supabase pgvector, and Google Gemini.
# Naija Tax AI 🇳🇬
## 🛠️ Tech Stack
A RAG-powered Nigerian tax AI agent. Users ask tax questions via Telegram and receive cited answers sourced exclusively from official Nigerian tax documents (NTA 2025, NTAA, NRSA, JRBA).
Built with n8n, Supabase pgvector, Google Gemini, and Telegram.
---
## Architecture
```
INGESTION PIPELINE (run once)
Google Drive (.md files)
↓
n8n Loop (one file at a time)
↓
Download → Extract Text → Split Out → Edit Fields (rename text → pageContent)
↓
Supabase Vector Store (insert)
├── Embeddings: gemini-embedding-001 (3072 dims)
├── Default Data Loader (JSON mode, {{ $json.pageContent }})
└── Recursive Character Text Splitter (chunk: 1500, overlap: 200)
QUERY PIPELINE (live)
Telegram Message
↓
AI Agent (Gemini 2.5 Flash)
├── Tool: Supabase Vector Store (retrieve-as-tool, Top K: 15)
│ └── Embeddings: gemini-embedding-001
└── Memory: Simple Memory (session = chat.id)
↓
Telegram Reply
```
---
## Stack
| Component | Service |
|---|---|
| Workflow automation | n8n self-hosted (Render) |
| Vector database | Supabase pgvector |
| Embeddings | Google Gemini (gemini-embedding-001) |
| LLM | Google Gemini 2.5 Flash |
| Document storage | Google Drive |
| Bot interface | Telegram |
| Document parsing | Docling (local CLI) |
---
## Prerequisites
- n8n self-hosted on Render (or any platform)
- Supabase project (free tier works)
- Google API key from aistudio.google.com
- Google Drive OAuth2 credentials
- Telegram bot token from @BotFather
---
## Step 1 — Supabase Setup
### Create documents table
Go to **Supabase → SQL Editor** and run:
```sql
create extension if not exists vector;
create table documents (
id bigserial primary key,
content text,
metadata jsonb,
embedding vector(3072)
);
```
### Create match_documents function
```sql
create or replace function match_documents (
query_embedding vector(3072),
match_count int default 5,
filter jsonb default '{}'
)
returns table (
id bigint,
content text,
metadata jsonb,
similar …