Building a knowledge-accumulation workflow for on-premises AI
Can an AI running on our own hardware gradually build a useful body of knowledge? I have built an initial workflow for finding and saving public documents, with Codex periodically reviewing the material and the research direction. Here is what the first tests established.
As the on-premises AI setup progressed, I wanted to build a way to accumulate useful knowledge after getting the models running.
The aim is to reduce the need to supply the same documents repeatedly and make previously checked findings useful in later work. Eventually, I would also like the system to study learning methods and propose what to learn next.
- DiscoverFollow official documentation links
- StoreKeep text and source metadata separately
- ReviewCodex checks original passages
- UseNormal-answer integration is still planned
Starting with knowledge that can be traced to its sources
This stage is about saving public documents and making source-checked notes available for later reference. It does not involve additional training that changes the model weights.
Each knowledge card records a concise summary, its source, a relevant passage, the retrieval time, its scope, and its limitations. Being able to return to the original material is central to the design.
Combining document discovery with periodic review by Codex
The initial collector follows documentation indexes and relevant links on selected official sites. Material that has merely been downloaded is kept separate from notes that have been checked against the source.
Codex is intended to review misinterpretations, outdated information, contradictions, and an excessive focus on one subject. I registered a task to run every six hours. At the point covered by this initial test report, its first scheduled execution had not yet been verified.
The planned subjects include academic and technical material, law, medicine, history, ethics, languages, mathematics, AI, IT, and current developments. Actual collection and test examples currently start with AI and IT. Sources for the other subjects will be checked and added gradually.
The first test: six sources and three source-checked notes
The initial experiment saved six public sources, including one documentation index. Three documents were discovered automatically through links in that index.
I registered three knowledge cards after reading the relevant original passages, then verified that they could be retrieved by a separate process. Connecting these notes to the business assistant’s normal answers is a later step.
I also ran 32 tests on the workstation and server. They cover behaviors such as withholding old notes when their source changes and rejecting a quotation that does not match its evidence. These are software checks, not a measure of overall factual accuracy or answer quality.
Recording a connection problem rather than treating it as success
The first attempts to fetch documents directly from the AI server timed out. Those attempts remain recorded as failures; they are not counted as successful document acquisition.
For now, the workstation collects public material and transfers the extracted text internally to the on-premises store. Text hashes are checked after transfer. Network routes and access restrictions were not changed.
Across two small workstation collection runs, there were nine HTTP requests and approximately 215 KB of response bodies. That measurement excludes the separate failed connection attempts from the server and protocol overhead such as headers.
Keeping collection small and reviewable
Each collection cycle is limited to three documents, with at least five seconds between requests and additional limits on request counts and response sizes. The collector checks site crawling rules and stops the cycle when it encounters a failure or restriction.
It does not bypass logins or authentication, probe non-public areas, or execute instructions or code found on a page. Public research is kept separate from processing internal business documents.
For law and medicine, the plan is to retain the date and scope of the material and avoid automatically applying it to an individual legal or medical decision. For history and ethics, differing interpretations should remain visible.
Next: testing whether the accumulated knowledge actually helps
The next implementation stage is to let the local AI summarize material and propose further subjects. It should also investigate newer learning approaches and explain why a proposal may help, how it could be tested, and what resources or costs it would require.
I then want to compare answers to previously unused tasks with and without the stored notes. Alongside the amount of material collected, the evaluation will look at practical benefits: less corrective work and evidence that is easier to follow.
There is something exciting about building this cycle of research, storage, use, and review on our own equipment. I will keep sharing the development process, including both useful results and things that do not work as expected.
This article describes the initial tests recorded on November 19, 2026. Fully autonomous learning, integration with normal business answers, and a demonstrated improvement in answer quality remain future implementation and evaluation work.
Public documentation consulted
Previous article: On-premises AI: from setup to everyday work
Development notes