request-intent-resolution

What problem does it solve?

Request intent resolution eliminates the guesswork of finding relevant prior support threads and turning historical resolutions into the correct next actions for a new intake message.

Core Features & Use Cases

• Thread-to-intent extraction (resolution-grounded): Extracts intent verb, objective, and an entity slot schema strictly from the resolution side to avoid intake leakage.
• Objective-space topology via embeddings + clustering: Builds a graph over embedded objectives, then uses community detection (prefer Leiden) to discover objective clusters and entity signatures.
• Soft cluster routing + diversified retrieval: Routes by a mixture of centroid similarity and seed evidence mass, retrieves within clusters, and applies MMR to reduce near-duplicate resolutions.
• Entity-aware slot elicitation: Computes which entity signature fields are missing from the known context to guide what to ask next.
• Retrieval quality evaluation with LLM-as-judge: Runs a continuous eval pipeline using complete/partial/irrelevant judgments, plus optimization via Optuna and experiment logging hooks.

Quick Start

Use the skill to resolve a new support message by extracting routing signals, clustering over the prior resolutions corpus, retrieving the most relevant historical resolutions, and returning missing entity slots for the next turn.