As the topic says I’m looking for recommendations on “where to start”. The goal is something like this:
I have a stack of pdfs and other document formats (mostly word & excel) which contain information about buildings, building codes and calculations. Everything is machine generated so it should be machine readable.
What “the AI” should be able to is tell me if a building code is still valid and used correctly (grab information from the internet from defined sources) and “how did we solve this problem X in earlier projects”
My current level of know how in regards to AI/LLM is almost zero so I am overwhelmed by the information that’s mostly out there because it mainly is two steps further in than I would need to.
Any hints, pointers and “how to”s are appreciated.
What you are describing is not one single technology problem or solution, its a whole collection of them lumped together and complicated by an ever shifting landscape of libraries, tools, and underlying tech stacks. And while I am a huge proponent of digital sovereignty, I have to ask: is there a reason this must remain offline and local?
To figure out what is possible, or works well/sucks, you might start with online services (say, o365 copilot if you have that) or tutorials doing this with existing “chat with your documents” offerings to compare capabilities and caveats. Solving the common problems of these major providers amplifies the barrier to entry doing it all yourself locally. If you absolutely cant upload your real documents, grab a sample set of ones that are not problematic and test with those. If the trillion dollar AI company cant do what you want, there are good odds its going to suck to try and solve for.
Consult Claude. Again, sounds like a weird solution but some idiot in a forum is only guessing at your skill level or confidence with various technologies to make any kind of meaningful recommendation. Work through options and considerations applicable to your specific use cases one at a time and go forth to experiment. If you decide to make your own, let the AI help you select and implement it. If something is new or unfamiliar to you, the bot probably has a better grasp of how to make it work. For the 100$ a year, that claude code subscription is a great entry level tech helper.
That doesn’t sound like content that will be easy to ingest.
AFAIK what you’re describing is not likely to work even slightly well on any AI local or not. According to companies that sell AI I’m sure it’s a breeze but in reality as of now ingesting “a stack” of various documents, parsing out building codes and verifying if they’re currently valid by running web searches is already a pretty tall order. Parsing out previous problem/solution pairs and correlating them with new problems adds another layer of complexity.
I imagine where you’d start would just be setting up an AI that can accurately parse building codes from your documents one document at a time. Then move on to one that can take one of these building codes and correctly verify if it’s currently valid… and so forth. At some point you may end up with something that’s mostly automated but the process won’t be anything like just pointing an AI to a bunch of documents in various formats and crafting a single prompt that accomplishes anything like what you’re describing.
You’re likely looking to just incorporate AI into an existing workflow. You’ll need to isolate a series of more manageable tasks and work towards developing each to a level of accuracy that’s acceptable. Depending on how each step goes you may end up with a fully automated system where one step seamlessly feeds into the next but it’ll be a process because there’s a lot that can go wrong at each stage.
I’d file that under “trade secrets” and “confidential documents”. Buildings that are to be checked and ingested by AI contain (but are not limited to) critical energy infrastructure buildings
Thanks for your reply, it’s a lot food for thought. Much appreciated
The term “PDF Text file” is misleading… PDF is one file format and text files use a vastly simpler format, just text. PDFs are a binary file format that includes support for many features allowing some PDFs to contain only images of text and others where the text is encrypted. Some are also just complex and difficult to parse.
Your specific PDFs may be relatively easy to read or not depending on how they were created. Even assuming you can easily extract all your files (PDF, word, excel, and other) getting AI to reliably parse out what you’re looking for can also be difficult… especially if you need high accuracy and your source documents vary.
It doesn’t follow that being machine generated makes files “machine readable”. Most types of “machine generated” files can only be easily read by software from the same company that invented the file format which is how the PDF file format started out. It was probably a good decade that PDF was a proprietary binary file format that Adobe held control over before they opened it up… similar story for word and excel file formats which also both started out as closed proprietary formats. PDFs may contain text but they are not simply text files so even just reading them can be problematic.
if by energy you mean NERC / CIP, then im not sure they care enough to restrict you. You could also make the information not CIP by censoring a few specific details. overall that is how we solve those kinds of problems.
however, if by energy you mean the kind involving lithium isotope separation and tritium production, you belong in the CUI COE discord to see what other people with UCNI are doing to either locally solve those kinds of problems or leverage an enclave in the cloud that is fully cmmc compliant.
I’m working on a personal project to build a RAG system that processes PDF files—mainly technical documents like datasheets, application notes, and programming manuals for STMicroelectronics microcontrollers. The goal is to store them in a SQLite database in a well-structured way so they can be searched efficiently.
I’ve learned that PDF files are not easy to parse. They’re excellent at defining how content should look on a screen or page, but extracting structured information from them is a completely different challenge.
Here are a few links to projects out there that will give you an idea of the obstacles you might encounter:
A short update for you kind folks who helped me so far.
What have I done up until now:
Installed a dedicated Linux machine on an old Dell T3600
Installed OpenClaw
configured Ollama
set up an email account to send and receive mail to interact with claw
So far the experience was… different. Some questions claw got right almost instantly (like “what is the part number for a lenovo E595 battery”) others he just ignored.
Then I told him to reply with “can’t do that dave” if he doesn’t know the answer and check emails every 15 minutes.
That worked but now I get “ollama usage alert” and I dialled it back to 60 minutes checks.
And yes, I am aware of the most recent CVE, applied the updates and I’m pretty sure claw hasn’t been infected. Mainly because he lives on a dedicated network and no interfaces have been made available to the outside. The only way to interact with it is shell, local web GUI or email.
Most disappointing job it did so far when I asked it today about the latest CVE (“What is „CVE-2026-33579“ and how do I fix it?”) where its answer was “I’m sorry Dave, I’m afraid I don’t know the answer to your question. For complex inquiries, a team member may follow up shortly.” But at the same time there was no notification like “a human needs to look at this”.
Where should I put my sights up next? What can I break now?
Thanks again to all who participated in this thread.
