Skip to content

ANZSIC Classifier

Classify any free-text occupation or business description into the correct Australian and New Zealand Standard Industrial Classification (ANZSIC) code — in under 3 seconds.

What it does

The ANZSIC Classifier takes unstructured text like "mobile mechanic" or "person who fixes cars at your home" and maps it to the precise 6-digit ANZSIC unit-group code that best describes that economic activity.

It works in two stages:

flowchart LR
    Q([User query]) --> R[Stage 1\nHybrid Retrieval]
    R -->|"20 candidates\nRRF-ranked"| L[Stage 2\nGemini Re-ranking]
    L -->|"Top 5\nwith reasons"| A([ANZSIC codes])

    style R fill:#1565c0,color:#fff
    style L fill:#6a1b9a,color:#fff
Stage What happens Technology
1 · Hybrid Retrieval Vector ANN search + FTS fused via Reciprocal Rank Fusion pgvector · PostgreSQL 15
2 · Re-ranking LLM reads candidates and selects best matches with reasons Vertex AI Gemini

Fast mode runs Stage 1 only (< 300 ms). High Fidelity mode adds Stage 2 (2–5 s) for production-quality results with natural-language explanations.

Key features

  • 5,236 ANZSIC codes indexed with 768-dimensional semantic embeddings (Vertex AI text-embedding-005)
  • Handles colloquial descriptions — "barista", "tradie", "runs a cafe" — that rule-based or FTS-only systems miss
  • CSV fallback — if Gemini returns no results, retries with the full code reference injected into the prompt
  • Hexagonal architecture — swap any component (DB, embedding model, LLM) by changing one line in container.py
  • 65 automated tests — 100% pass, no live services needed for unit tests

Quick navigation

Tech stack

Component Technology Version
Language Python 3.12+
Embeddings Vertex AI text-embedding-005 768-dim
LLM Vertex AI Gemini gemini-2.5-flash
Vector DB PostgreSQL + pgvector 15 + 0.8.0
UI Streamlit 1.35+
Models Pydantic v2
Tests pytest 8.0+