Internal Document Retrieval

Decision question

Analysts at a B2B media company repeatedly searched internal documents and earlier reports to answer client questions. The company wanted to reduce that search time without returning answers detached from its own research.

Records used

The system used the client's internal document collection and recurring analyst questions. The documents were split into retrievable passages and indexed in a vector store.

Work performed

I built a retrieval layer that accepted a question in plain English, ranked document passages by relevance, and supplied the selected passages to a generative interface. The user could review the retrieved material instead of receiving an unsupported answer from the language model.

The useful system had separate parts: document ingestion, passage retrieval, language-model drafting, and user review. Calling the entire chain "AI" would hide the operation that mattered most, which was finding the correct internal evidence.

Observed result

After adoption, analysts reported spending about 40 percent less time finding prior research. The client retained the tool as a recurring paid product.

Business use

The interface reduced repeated manual search and made earlier work available during new analysis. It also turned a one-time build into a tool the consulting firm could continue supporting.

What the result does not establish

The time reduction came from an operational comparison of analyst workflow before and after adoption. It was not a controlled time-and-motion study. The result does not measure answer quality by itself. Retrieval accuracy, source coverage, and user review remain separate checks.