Show HN: Emotion-Weighting API with GPT-5 (lightcapai.medium.com)

A developer demonstrated an emotion-weighting API that uses GPT-5’s function-calling to integrate a custom emotion-analysis tool into the model’s response flow. The pattern is two-step: send a prompt plus a JSON-schema-defined function (e.g., analyze_emotions), let GPT-5 return a function_call, execute that function in Python (rule-based or lexicon scorer returning JSON weights like {"joy":0.7,"sadness":0.2}), then append a function_call_output (tied to the call_id) and call GPT-5 again to produce a grounded final answer. The example is wrapped in a FastAPI endpoint that accepts a prompt, forwards tools to OpenAI’s SDK (client.responses.create with tools), handles function_call items in response.output, and returns both the numeric weights and the model’s interpretation. This approach is significant because it shows a pragmatic way to reduce hallucination and make LLM outputs auditable and reproducible by anchoring them to deterministic tooling. Key technical details: define strict JSON schemas so GPT-5 emits well-formed calls (strict:true and additionalProperties:false), inspect response.output for type=="function_call", parse arguments, execute the tool, and include a function_call_output message before the final model invocation. Best practices include limiting available functions, validating calls and parameters, using a system prompt to restrict invented tooling, and protecting API keys and external data sources.