Procurement Code
Search Bot
Telegram bot + REST API for searching government procurement classification codes. Built with vector search, RAG architecture, and LLM reranking — high-accuracy results over a 20,000+ item catalog.
20K+
Codes indexed
80%
Top-1 accuracy
3–5s
Avg response
100
Batch API capacity
// context
Problem & Solution
❌ Problem
- Procurement specialists manually browse a 20,000+ item classifier
- Commercial product names don't match official classifier terminology
- Errors lead to rejected bids and legal penalties
- No existing tool combined classifier search with real procurement registry signals
✅ Solution
- Natural language search — type product name, get classification codes
- Hybrid search: vector similarity + full-text + procurement signal via RRF
- LLM reranking to surface the most relevant code first
- REST API for direct integration with ERP systems
// architecture
System Architecture
Telegram Bot
aiogram 3.x
User input
→
searcher.py
Search pipeline
Core logic
→
Supabase
PostgreSQL + pgvector
Vector DB
→
OpenAI
GPT-4o-mini + Embeddings
LLM layer
→
FastAPI
REST API / ERP
ERP integration
// search pipeline
1
Verdict Detection SQL Tool
Parallel registry lookup + full-text search count. Determines path:
specific / ambiguous / classifier_search2
HyDE Query Expansion GPT-4o-mini
Generates a hypothetical classifier-style description to improve embedding quality. Temperature = 0 for consistency.
3
Hybrid Search — RRF Fusion SQL RPC
Reciprocal Rank Fusion combines: vector search (×1.0) + full-text on name+enriched (×2.0) + registry signal (×1.0). k=60, 40 candidates.
4
LLM Reranking GPT-4o-mini
GPT-4o-mini selects top-10 from 40 candidates, filters parent/child duplicates, prioritizes leaf-level codes.
// demo
Bot in Action
// features
Key Features
Hybrid Vector Search
RRF fusion of vector similarity, full-text search, and procurement registry signal. Handles commercial names, abbreviations, and typos.
Smart Caching
MD5-based query cache in Supabase. Repeated queries return instantly (0.5s vs 3s). Auto-expires after 30 days.
REST API + Batch
FastAPI endpoint with single and batch search (up to 100 queries). Swagger docs auto-generated. Direct ERP integration.
Registry Signal Boosting
Uses real procurement registry data as a ranking signal to prioritize codes that are actually used in contracts.
Feedback Loop
Every user selection saved to database. Enables future fine-tuning and quality benchmarking against specialist choices.
Data Enrichment
Classifier codes enriched with GPT-generated synonyms. Registry data loaded from official XML source (55K+ records).
// results
Metrics & Results
80%
Top-1 accuracy
85%
Category accuracy
~3s
Avg response time
20K+
Codes indexed
0.5s
Cached response
100
Batch API capacity
// tech stack
Built With
Python 3.11
FastAPI
OpenAI API
Supabase
pgvector
aiogram 3.x
Railway