Ambiguity Resolver for LLM-based systems

[ShaliniAnandaPhD]

I just open-sourced a new microservice inside Neuron that targets a specific but critical failure pattern in AI systems:

When polite, indirect language hides urgency — and your LLM takes it literally.

What It Does

• ToneAgent: Detects politeness markers (hedges, subjunctives, minimizers, apologies)
• IntentResolver: Extracts the true user intent (e.g. billing issue, account access)
• UrgencyScorer: Calculates realistic urgency — even when masked
• OutputAgent: Logs everything to structured JSON with UUIDs and timestamps

Features

• Command-line and FastAPI support
• YAML-configurable thresholds
• Labeled evaluation dataset (evaluation_dataset.json)
• GitHub Actions CI with test coverage
• Dockerfile included
• Structured log output
• Example integration with a support ticketing system

Sample Query It Handles

“Sorry to bother you, but I can’t log in and I have a meeting in 10 minutes.”

Most systems would treat this as low priority. This microservice correctly flags it as high urgency.

I’m Looking For:

• Ideas for edge cases to test
• Suggestions for improvements or new detection patterns
• Feedback on multilingual support or cultural tone variance
• Thoughts on how to open it up for public testing (Hugging Face, Binder, REST API, etc.)
• Contributors or testers interested in reliability tooling for LLMs

GitHub repo: https://github.com/ShaliniAnandaPhD/Neuron/tree/main/Microservices/ambiguity

Happy to answer questions or collaborate — let’s build more context-aware systems that don’t fail on politeness.