Juq 699 -

| ID | Requirement | Details | |----|--------------|----------| | | Real‑time Tag Suggestion | • Triggered after 3 characters in title/description. • Uses a lightweight transformer (e.g., DistilBERT) fine‑tuned on the organization’s corpus. • Returns up to 5 ranked tags with confidence scores. | | FR‑02 | Tag Acceptance UI | • Checkboxes next to each suggestion. • “Add all”, “Reject all”, and “Edit manually” actions. • Persist accepted tags in the item’s metadata. | | FR‑03 | Contextual Search API | • New endpoint POST /search/contextual that accepts free‑text queries. • Internally expands the query using the same model + synonym dictionary. • Returns ranked results with relevance score and highlighted matches. | | FR‑04 | Admin Settings Page | • Slider for confidence threshold (0‑1). • Multi‑select for allowed tag vocabularies (e.g., “Finance”, “HR”). • Toggle for “Auto‑apply tags without confirmation”. | | FR‑05 | Analytics Dashboard | • Charts: Tag suggestion acceptance rate, avg. confidence of accepted tags, search success (click‑through rate). • Export CSV/JSON. | | FR‑06 | Fallback & Privacy | • If the model fails (e.g., latency > 500 ms) fall back to keyword‑only suggestions. • No personal data leaves the organization; the model runs on‑prem or in a VPC‑isolated cloud environment. |

Users spend a lot of time manually classifying and locating content (documents, images, tickets, etc.). Existing keyword‑only search returns too many irrelevant results, while the current tagging system is static and requires users to remember a predefined taxonomy. juq 699