How Search & Retrieval Works

Document Search

Sevrel uses a technique called Retrieval-Augmented Generation (RAG). Unlike a general-purpose chatbot that relies on pre-trained knowledge, RAG supplies relevant passages from your documents as source evidence. This can reduce unsupported answers, but results can still be incomplete or wrong, so verify material claims.

What Documents Does Sevrel Understand?

Sevrel can read and analyze most document formats used in commercial real estate workflows:

How to Get Better Search Results

The quality of Sevrel's answers depends on how well it can find the right documents. Here are strategies to improve results:

• •Name the property. “What is the base rent at Maple Crossing Plaza?” is much more effective than “What is the base rent?”
• •Mention the document type. “Check the 2025 shared cost adjustment for Cedar Point Center” helps Sevrel narrow its search.
• •Include dates or time periods. Financial data often exists across multiple years. Specifying “Q4 2025” or “2024 budget” reduces ambiguity.
• •Use tenant names. When asking about a specific lease or tenant, include their name exactly as it appears in your documents.

Real-Time Thinking Indicator

While Sevrel processes your query, a live timeline shows exactly what it's doing at each step — searching for documents, reading specific files, and synthesizing an answer. This transparency lets you see that Sevrel is working with real data, not guessing.

This also helps with troubleshooting — if Sevrel searched the wrong documents, you can rephrase your query with more specific property or tenant names.

Indexing and Retrieval

Sevrel maintains a search index over your connected document library. A background crawl reads the files it can reach — up to a per-run file limit, and files above 50 MB are skipped — extracts their text, splits it into passages, and stores those passages encrypted, so most questions are answered from the index rather than by re-downloading your files. Where semantic search is enabled for your organization, each passage also gets a vector embedding.

At query time, Sevrel runs several retrieval passes and merges the results: keyword matching over the index, semantic matching over the same index when those embeddings are available, and a live search against your connected storage provider to catch files that have not been indexed yet. Only the highest-ranked passages are sent to the model, rather than your whole library. The number of retrieval and reasoning steps varies with the request, available evidence, and configured limits.

Availability depends on successful provider discovery and indexing. New or changed files may take time to become searchable, and users should confirm that the intended source was retrieved. See Privacy & Data Isolation for what Sevrel stores.

Understanding Source Citations

Responses based on retrieved documents are designed to include numbered citations linking back to source evidence. Citations can be incomplete or mismatched, so verify material claims before relying on them.