Summary
SolidKnowledge turns a CAD archive from files we store into engineering knowledge we can actually use. I developed the prototype during my internship in the Innovation Lab at Dimensional Innovations, where twenty years of SolidWorks projects sat on a network drive with no memory of what they contained.
This page follows how we made previous engineering discoverable through evidence, search, and continuing care. A short clip of the search opens it, and a full recording of the prototype closes it.
CAD knowledge graph Evidence-first ingestion Semantic search Supervisor agent Human-in-the-loop
Picture the people who spend their days designing and fabricating: every finished project leaves behind details that worked, problems solved the hard way, and approaches proven in the field. When a new job arrives, a shape or connection often feels familiar, even if nobody remembers where it came from. The same quiet question comes up: have we built this before?
Finding that earlier solution can be harder than redrawing it, because remembering a detail rarely means remembering its filename or project folder. We took on that gap during my summer 2026 internship at Dimensional Innovations, asking how the archive could recover the engineering people had already worked out. Our starting point was understanding what twenty years of delivery had left behind.
The archive holds +360,000 SolidWorks files across about twenty years and hundreds of projects, with parts, assemblies, and drawings gathered into project folders. Together they preserve the accumulated decisions of people who found ways to make things work. Their value lies in the engineering inside them: a reliable connection, a material choice, or a fabrication approach worth using again.
Those folders serve the practical work of delivering projects and managing files, but their structure offers few clues about which design could help with today’s problem. A file system tells us where something lives; filename search still asks us to remember its name. We wanted to begin with what someone needed, which raised another question: what if the archive could explain itself?
We developed SolidKnowledge to make that stored engineering understandable and findable, with a route back to the original work. The map above connects evidence gathering and knowledge reconstruction to a semantic space, where descriptions with related meanings sit near one another so people can discover and reuse previous designs. Making those connections starts with understanding what a single CAD file contains.
No single source explains a design completely: geometry supplies measurements, drawings preserve an engineer’s material notes, and the project folder gives the work a setting. We set a rule that these facts must come before interpretation. Reading them together gives the system a basis for describing the object without expecting its filename to carry the whole engineering story.
Ingestion, the process of reading and gathering those sources, runs on a workstation with SolidWorks and assembles an evidence card for each file: everything we can verify about the object before any interpretation. It brings measured geometry and authored text together with four rendered views, giving the later interpretation something concrete to work from. We wanted that evidence to remain useful when a person, rather than a machine, encountered the result.
Evidence first. Interpretation second.
A name, size, and date tell someone that a file exists, but leave much of its engineering hidden until they open it. An evidence card holds what we can actually verify about the design: its shape, material, and project, with four consistent views so a reader can recognize the object before deciding whether to investigate the original CAD.
Every field leads back to evidence, so a dimension rests on measured geometry and a material rests on authored drawing text rather than a guess about a filename. A file tells you it exists; an evidence card tells you what we actually know about it. To give that card enough context, we needed the relationships connecting the object to the engineering around it.
A knowledge graph maps engineering objects and the relationships between them: assemblies contain parts, and drawings document models. A shared component can belong to two parent assemblies even though its file occupies one place in the folders. That distinction matters when we want to understand how a detail was used, rather than simply locate the file that stores it.
We reconstructed references and cross-verified them in two independent ways: through the SolidWorks Document Manager API, an interface for reading file information, and by opening files in SolidWorks. Those connections let a file draw context from its parent assembly and the drawings that document it. The evidence card now reaches beyond the individual object, but it is still unstructured: verified facts waiting for interpretation.
An AI reads the evidence card and writes the knowledge card: the same object, now structured and interpreted, ready to search. One half explains the object in engineering language through its summary, design intent, materials, geometry pattern, and project context. The other half makes that meaning usable for retrieval through a controlled object type, tags from DI’s vocabulary, and search aliases reflecting the words people type.
A reasoning field records the evidence behind each claim, keeping the knowledge card connected to the drawing text or assembly context that supports it. Materials come from authored notes, rather than a guess based on appearance. The knowledge card’s text then becomes an embedding, a numerical representation of meaning that lets the system compare descriptions, although one description cannot emphasize every aspect equally.
The same design can matter because of its shape, its material, or the purpose it served, depending on the problem someone is trying to solve. A single summary tends to emphasize one of those reasons at the expense of others. We wanted a useful precedent to remain discoverable when the next person’s question approached it from a different direction.
We therefore describe each evidence card from five perspectives: what it was for, what shape it is, what it is made of, where it lived, and the words someone would type. Each perspective becomes its own searchable representation, and together they form the knowledge card that the semantic space stores, so discovery still leads to one engineering object with its evidence. With those representations in place, someone can begin a search with the detail they actually remember.
Consider a designer looking for a curved fascia, a facing panel hung on a concealed cleat, who remembers the arrangement but none of the project details. They type “curved fabricated assembly with concealed support” into search. The sentence enters the same semantic space as the archive’s descriptions, letting its meaning connect with previous work regardless of how the files were named.
In this example, the answer is yes: a fascia from a 2016 stadium job appears with its knowledge card, including AL 5052 aluminum and powder coat. The match explanation helps the designer judge its relevance before one click reveals the source file. That makes earlier work available for examination, but its usefulness depends on the card still describing the source accurately.
An assembly can change months after delivery, leaving a knowledge card that was accurate in March describing an old version by July. The words remain just as convincing, which makes stale knowledge difficult for a reader to recognize. We designed the automation to keep published knowledge aligned with its sources, dividing the work according to the resources it needs.
Full-archive extraction takes roughly one to two weeks per workstation because SolidWorks is the bottleneck. We scheduled the Night Worker, the extraction process, to gather evidence when Windows workstations with SolidWorks seats are free. The Enrich Worker, which turns that evidence into searchable knowledge, needs no SolidWorks seat and runs around the clock, connected to extraction by a queue.
We gave the Supervisor Agent a restrained role: observe the workers and compare the live archive with published knowledge, project by project. It reports changes and proposes action, but never touches source CAD or initiates repairs independently; re-enrichment requires human approval. Together, these roles keep responsibility clear as the memory changes, with people deciding how to respond to its findings.
The workers build the memory. The Supervisor keeps it true.
SolidKnowledge connects the question someone can ask today with engineering that might otherwise remain buried in an old project folder. The evidence and relationships behind a knowledge card help that person examine what a previous design offers, while the workers and Supervisor maintain the connection to its source. We developed those parts together because discovery depends on knowledge worth returning to.
For the people designing and fabricating the next project, “have we built this before?” can now lead to a design they can inspect and a source they can open. We wanted the experience of earlier work to remain available to the people who need it again. The recording below brings that idea into view through the prototype on the real archive.
The recording is a screen capture of the working prototype on the real archive. Watch for how a designer’s sentence turns into ranked past designs, how a knowledge card lets them judge a result before opening CAD, and how the operator’s side of the tool keeps the memory current.
If it doesn’t load, use the direct link below.
Useful engineering stays difficult to recover when discovery depends on filenames and someone remembering the project.