AI Studio
Bring your own documents. Start asking questions in minutes.
The idea was simple. Bring your own documents, connect a drive or upload a ZIP, and start asking questions. Getting that to work in minutes was the hard part.
At Fulcrum, we were building this for insurance, ecommerce & enterprise teams. Every customer brought their own documents. The main challenge quickly became ingestion and indexing at scale.
How much extraction does this document need?#
We brainstormed different approaches. Using the strongest model for everything would mean spending that effort even on documents where simple OCR could do the job.
But simple extraction wasn’t enough for everything either. Images and complicated layouts needed more help.
We designed a smart router to make that choice. It could use simple OCR, a hybrid approach with a vision-language model (VLM) for images and simpler extraction for text, or full VLM-based OCR.
- Drive link / ZIPThe customer's documents
- Smart routerChoose extraction capability
That gave us a way to handle different documents without treating every page as equally difficult.
Don’t make the first question wait for everything#
Extraction was only one part of the wait. We still had to chunk the content and index it before retrieval could be useful.
We used hybrid chunking, with lazy indexing around it. First, we extracted the structure of the document and its high-level details. We didn’t have to finish chunking the entire collection before someone could ask a question.
The question itself helped decide what to do next. We identified the relevant sections and prioritized their indexing.
- Ask a questionFind the relevant section
- Prioritize itChunk & index that section
- Retrieve more detailAnswer with richer context
Some questions could be answered with the initial information. Answers improved as more of the document was indexed.
There was a tradeoff here. The system’s capability grew as indexing progressed. Some early answers were basic, then became more detailed as the relevant content became available.
Finding the answer#
Once indexed, we combined dense retrieval and keyword search in Qdrant, merging the results with reciprocal rank fusion. We also routed queries between document retrieval, SQL and tools, depending on what the question needed.
The agent used that context to produce answers with citations. The retrieval engine had to do more than find something that sounded related. It had to help the customer get back to the source.
What changed for the customer#
They could configure a drive link or bring a ZIP of documents and start chatting in minutes.
The decision that made the biggest difference was when to do the work. Understand the collection first, then let the questions help prioritize the detail.