Legal Tech
Answer Suggestion for Legal Interrogatories
A semantic search layer that finds the closest previously-answered question in a litigation knowledge base, so legal teams answer interrogatories by review rather than from scratch.
The problem
Interrogatories are the formal written questions one side of a litigation sends the other. In practice, companies answer similar questions the same way every time — they disclose a consistent amount of information regardless of which document the question arrives in.
That consistency was being maintained manually. Someone had to remember, or find, how a comparable question had been answered before, across a knowledge base of prior litigation documents. The client, a US legal-tech SaaS platform, wanted that lookup automated inside the product they already shipped.
What I built
An answer suggestion system built on document similarity. A semantic layer matches an incoming question against the knowledge base and surfaces the closest prior questions along with how they were answered.
Pure semantic similarity was not accurate enough on its own, so I added custom NER models that tag the knowledge base with domain metadata — entities specific to the legal documents in play. That metadata then drives faceted search, letting the matching narrow by attributes rather than relying on embedding distance alone.
The whole thing had to fit into an existing production system, so it was built as a layer the client’s platform could call rather than a separate application.
The outcome
- Sharply reduced turnaround time on interrogatory responses.
- Sharply reduced the cost associated with servicing each request.
- Legal teams moved from drafting to reviewing — the expensive part of the work became a judgement call on a suggestion rather than a search.