Searching Prior Research in Plain English

Analysts kept answering questions someone had already answered in an old report. The assistant found that report.

Client: A B2B media company's marketing and PR department, while I was at Ringer Sciences.
My role: I designed and built the search assistant end to end.
Result: About 40% less research time over a one-quarter pilot, and the client kept it as a paid product.
How it was measured: The department's own estimate at the end of the pilot.

Result: analysts spent about 40% less time on research, and the client kept the assistant as a recurring paid product.

The problem

While I was at Ringer Sciences, analysts at a B2B media client spent hours digging through internal documents and old reports to answer the same kinds of client questions. The answers existed. Finding them was the slow part.

What I built

I designed and built a retrieval assistant end to end. An analyst types a question in plain English, and the assistant answers from the company's own documents and shows the passages it used.

Finding the passages. I split the documents into short passages and chose how they were indexed and searched. Each question pulled the most relevant passages, and a second pass put the best ones first before anything reached the language model.
Serving. A small server passed each question between the search interface, the passage index, and the model.
Drafting. The language model wrote its answer only from the passages it was given, and the analyst could open those passages to check it.

The part that mattered most was finding the right internal passage. Calling the whole chain "AI" hides that.

How the client used it

The client's whole marketing and PR department used the assistant through a one-quarter pilot. Over that quarter the department reported about 40% less time spent on research, mostly because people no longer hunted for citations themselves. The figure is the client's own estimate, since its internal processes weren't shared with us.

Analysts stopped repeating searches and started building on earlier work. The client kept the assistant after the pilot, so for the consultancy a one-time build became recurring paid work.

Built with (for technical readers)

OpenAI embeddings stored in Pinecone, retrieval of the top passages followed by reranking, a FastAPI server, and a language model that drafts only from retrieved text.