A mini “tool-using” chatbot implemented with rule-based routing. The system decides when to call external tools, extracts structured tool inputs from natural language, and then formats raw tool outputs into concise responses using a small open-source LLM.

• GitHub: TODO (paste repo link)
• Demo: TODO (optional)

Notes

What I built

• Safe Math Tool (AST evaluator): Parses expressions with Python ast and evaluates only whitelisted numeric operations to avoid eval-style security risks.
• Maps / Geocoding Tool: Calls OpenStreetMap Nominatim to convert place names into latitude/longitude, with graceful handling of missing results and API errors.
• File Search Tool: Searches a local repo (e.g., SciPy) for filenames matching a query using os.walk, returning top-k results efficiently.
• Tool Registry: A small registry abstraction to register and retrieve tools by name (extensible design).
• Dispatcher + Argument Extraction: Keyword routing plus regex parsing to pass clean arguments (e.g., extracting "Paris" from “Find the coordinates of Paris”).
• Toolformer Wrapper: A single toolformer_answer() function that (1) dispatches to tools and (2) uses a mini LLM to rewrite tool outputs into a clear, user-facing response.

Key takeaways

• Tool-using assistants require more than “calling an API”: accurate routing + clean argument extraction matter as much as the tools themselves.
• Security and robustness are first-class concerns (e.g., AST-based evaluation instead of unsafe execution).
• Even small LLMs can be useful for formatting and response polish, and practical optimizations like pipeline caching significantly improve usability.

Tech stack

Python, AST, Regex, requests, OpenStreetMap (Nominatim), HuggingFace Transformers (google/flan-t5-small)