Personal project using Python, Ollama, and Llama 3.2. The shop inventory is included with each query.
The challenge
The default context window silently truncated the inventory. Increasing it to 32,768 tokens let the model receive the full catalogue. UTF-8 output handling also fixed Windows console crashes.
Results
- Natural-language lookup against roughly 1,465 inventory entries.
- Runs locally without API keys or cloud requests.
- Inventory instructions and lower-temperature sampling reduce invented recommendations; they do not guarantee correctness.