A prototype fails a test. Someone changes the material, writes a note, and updates the drawing. Months later, a colleague asks why the old design was rejected. The answer may be split between a test report, a meeting record, and a file with an unhelpful name.
That is a useful question for a retrieval-augmented generation system: find the evidence behind a decision, then explain it with references. It is a different job from asking a general chatbot to suggest a design. We build RAG AI systems around an organization’s own material, so people can inspect what supports an answer.
Burlington, Vermont has a concrete setting for this kind of work. UVM Innovations supports research commercialization. Generator provides prototyping facilities in the South End. VCET brings startups, freelancers, and remote technology workers together in downtown coworking space. They are mentioned here as local background, not as examples of our work. The examples below concern hypothetical businesses.
Keep research findings separate from product decisions
UVM’s startup pathway describes technology development and pilot projects, with market feedback and real-world testing helping refine the technology. It also points founders toward Generator and VCET. That combination of research, physical development, and business support gives Burlington founders several routes from an experiment toward a product.
Consider an independent product startup with its own test reports and customer interview notes. A question such as “Why did we change the enclosure?” might need both kinds of evidence. The test report could describe a failure; the interview note might explain why a proposed replacement was inconvenient to use. A useful answer should identify those as different observations rather than turn them into one authoritative conclusion.
Preserve the project name, author, date, and document type when preparing the search library. A preliminary finding should still look preliminary when it appears in an answer. If a later decision accepts, rejects, or qualifies that finding, the system needs to retrieve that relationship too.
This does not make the assistant an engineer or a product manager. It gives the person making the decision a shorter path to the underlying work.
Search prototypes without confusing them with production
Generator’s Burlington facilities include wood and metal shops, an electronics lab, laser cutters, and 3D printing. Its description of 3D printing includes rapid prototyping. That is a specific local reminder that product knowledge can include physical iterations, not just a polished manual.
For a hypothetical small design business, the search library might contain assembly notes for several prototypes alongside a released product guide. A question about a fastener or component should not return an obsolete experiment merely because its wording is a close match.
Treat product identifiers and revision labels as search constraints. Ask which version the person means when that information is missing. Label an experimental note as an experiment, and make the current approved document easy to recognize. The assistant should expose disagreement between sources instead of smoothing it away.
Drawings and photographs need care as well. If the answer depends on a dimension shown only in an image, a text-only index may not contain it. Keep the original file accessible and explain what the search actually covers. Do not convert a missing measurement into an estimated one.
A retrieved setup note also cannot replace equipment instruction, supervision, or a responsible person’s approval. For work with physical consequences, the useful feature is finding the applicable record, not issuing confident operating directions.
Make a handoff searchable without sharing every file
VCET’s downtown Burlington coworking community includes startups, freelancers, and remote technology workers, with business advice and meeting spaces. For an independent team assembling people with different specialties, a practical question is how much information each collaborator needs.
A hypothetical software startup bringing in a contract designer could provide an approved project library: the product brief, decisions about user flows, and the relevant interface requirements. That person should not automatically gain access to unrelated customer records or internal financial material.
A RAG interface needs the same boundaries as the documents behind it. Permissions should determine which passages can be retrieved before an answer is generated. Hiding a link afterward is insufficient if the answer already includes restricted text.
Plan for the end of the handoff too. When a collaborator’s access is removed, test that old conversations, saved answers, and source links do not preserve access the team intended to revoke. The right controls depend on the system, but they belong in the scope before a shared search tool is introduced.
Build the first test around one completed product change
Choose a question that has a documented answer and someone who can judge it. For a prototype-heavy team, one completed design change can be more informative than importing every shared folder.
A compact evaluation could cover:
- The reason a component changed, with the correct test report attached.
- Two similar product names whose documents must remain separate.
- An experimental result that was later rejected.
- A question requiring a drawing the text index cannot read.
- A request for a file the user is not allowed to access.
- A question for which no supporting record exists.
Inspect both the retrieved material and the written answer. A good summary of the wrong revision still fails. Record unsupported claims separately from missed documents, and compare the result with the search tools the team already has.
Our AMC Boston Doc Chat case study shows a system we built for the Boston Chapter of the Appalachian Mountain Club, with answers linked to supporting passages. It demonstrates the retrieval approach, not results for a Burlington manufacturer. Our account of searching historical archives explains why known-answer tests matter when judging a new collection.
Decide whether retrieval is worth maintaining
A small product library may need clearer filenames and an owner for each document, not custom AI. If nobody knows which specification is current, indexing the folder will not settle the disagreement.
Custom retrieval becomes worth evaluating when repeated questions cross several document types, the collection changes, and staff need a checkable explanation rather than a list of matching filenames. Someone must still maintain source material and review failures after launch.
For Burlington teams deciding where to begin, distinguish searching from reviewing everything. Finding the record behind a product decision suits RAG. Extracting the same fields from every test report is closer to AI document analysis. Getting that distinction right keeps the first project focused on the actual work.