Implement tool / function calling

Transaction-support bot that answers from tools, not memory

Build a support chatbot that answers questions about payment transactions (status, refund eligibility, dispute filing) by calling four tools against a local transactions database instead of guessing. You create the data yourself: a short Python generator seeds a SQLite table with ~200 synthetic transactions covering settled, failed, pending and already-refunded cases, so no real payment system or merchant account is involved. Validate every tool's arguments with Pydantic before execution and feed structured results back into the conversation. Put a confirmation turn in front of the one tool that mutates data.

A SQLite transactions table you seed yourself — ~200 synthetic records (txn id, amount, status, timestamp, refund window) from a 30-line Python generator

1. Seed the synthetic transactions DB and write the three read-only tools · ~1.5h
2. Wire the tool-call loop: parse tool_calls, dispatch, feed results back into the turn · ~2.5h
3. Add Pydantic argument validation and a confirmation turn before the dispute tool · ~1.5h
4. Log 10 sample conversations and write the README with the tool schemas · ~1h

• At least 3 distinct tools defined with JSON-schema parameters and docstrings
• Pydantic models validate and reject malformed tool arguments before execution
• A destructive tool (raise dispute) requires an explicit user confirmation turn before executing
• 10 sample conversations logged showing correct tool selection, including one where the model asks a clarifying question instead of guessing

Evidence a recruiter can check

• The four tool schemas beside a transcript log of 10 conversations, annotated with which tool fired for each intent
• A transcript where the model asks a clarifying question instead of inventing a transaction id
• Unit tests that feed malformed arguments to each tool and assert they are rejected before anything executes
• A recorded run of the dispute tool refusing to mutate a row until the user confirms in a separate turn

Built a transaction-support bot on LLM tool calling — four Pydantic-validated tools over a transactions database, with a confirmation gate on the one destructive action, so every answer traces back to a row instead of the model's guess.