A stack you cannot draw is a stack you cannot budget. Long before price comes up, executives evaluating Iternal reach for a pen: are we deploying Blockify, or just a general model? Does Blockify even belong in an AirgapAI conversation? Sequence is the architecture question buyers raise more than any other, and it fits on a napkin.
Product Architecture
Which Components Do You Need, and
In What Order Do They Run?
Three parts and one sequence: the ingestion pipeline that structures your documents, the data set it produces, and the local assistant that answers against it — plus what each part is not.
Built from real buyer questions in our sales meetings
Direct Answer
Three parts, one sequence. An ingestion pipeline turns your source documents into a structured data set; that data set is loaded into the local assistant; the assistant answers against it using a language model running on the device or on a server you own. Iternal names those parts Blockify, the data set, and AirgapAI. AI Assist is a commercial package rather than a box in the diagram: Iternal and Dell built it as a Dell-exclusive bundle of AirgapAI software, white-glove deployment consulting, AI Academy training and enterprise support.
The limit: the pipeline only improves the data it touches. Iternal says it plainly — Blockify is no silver bullet. Getting the order right does nothing for a source document you left out of the run, and no arrangement of these three parts compensates for an incomplete corpus. Which material goes in is your decision, made before any of this software executes.
Each part runs without the others, and the accuracy claim needs two of them. Run the pipeline alone and feed the result to AI you already own: Iternal builds Blockify architecture agnostic, deployable independently of any hardware platform, and its processed data set can be consumed outside AirgapAI by any application. Run the assistant alone with no data set switched on and it still answers — from the loaded language model only. The accuracy Iternal claims belongs to the pairing.
What you deploy and what you buy are separate drawings. A lightweight Blockify utility ships on the device with AirgapAI; a server-side deployment of the same process handles enterprise volume on your own hardware or in a tenant you control. Choose the scale from how often you rebuild data sets, and settle that licensing in writing. For more information on what the pipeline accepts, visit the input-formats page.
The Order, End to End: Source Document to Answer
Five stages, named once and used consistently below. Every arrow is a hand-off you can inspect or replace.
01
Source documents
Material you already own: manuals, contracts, policies, reports.
02
Ingestion pipeline
Blockify reads the text, distills it, attaches metadata and tags.
03
The data set
IdeaBlocks: small units, each carrying a critical question and a trusted answer.
04
Local assistant
AirgapAI loads the data set and retrieves across the blocks you switched on.
05
The answer
Written by the loaded model, with the blocks it leaned on shown beside it.
The model sits under stage four, not inside it. AirgapAI is model agnostic and takes a language model you supply; the model setting carries general knowledge plus tone, while your data set carries the facts. Users switch between a local model and a server model as the work demands, and Iternal states the model can be swapped without disturbing the corpus — keeping model choice and data preparation independent decisions.
What Buyers Ask Before They Ask About Price
The same boundary questions arrive once the diagram is on the table. Five of them, in buyers' own words, with the answer Iternal gives:
| The question, as buyers put it | The answer |
|---|---|
| Are we deploying Blockify, or just a general model? | Your call. The assistant runs on a model alone; the pipeline turns your material into a data set it can cite. |
| Are we talking about AirgapAI only, or does Blockify belong here? | AirgapAI runs on the AI PC. Blockify is built for the enterprise sale, deployed on your servers or in a tenant you control. |
| Do you use Blockify to ingest the files that get dragged in? | Yes. Dragged-in content goes through the Blockify process and appears as a data set in the app. |
| Are IdeaBlocks made from Blockify? | Yes, that is one way to make them. |
| Do you rely on other components in the stack to tune the result? | Surrounding components stay in place and can be swapped. Blockify slots between your extraction step and your storage and search. |
One further question decides the commercial shape: are these finished AI products, accelerators, or a platform we can build on? All three, Iternal answers, depending on how much ownership you want — from licensing the foundational technology to a turnkey service. Settle it early: it names who runs stages two and three.
Each Part Runs Without the Others
A sequence is not a bundle. The three parts hold a fixed order together and each stands alone, which makes the architecture buyable in stages.
- The pipeline, alone. Iternal designed Blockify architecture agnostic: it drops into an existing stack without changing the vector database, the re-ranker or the extraction step already in place.
- The data set, elsewhere. Iternal describes the output as assistant agnostic: the cleaned result can be consumed outside AirgapAI by any application, including one your developers write.
- The hardware, your choice. Iternal is agnostic on hardware platforms and silicon, working across Intel, NVIDIA and AMD, and Blockify installs on the infrastructure you already run rather than requiring one of its own.
- The assistant, unaccompanied. AirgapAI answers with no data set active, on the loaded model alone: useful for drafting, silent on your organization.
Two scales of the same pipeline, and both are real. A lightweight Blockify utility ships on the device with AirgapAI, reached through the my-data control, for drag-and-drop packaging. A server-side deployment runs the same process at enterprise volume, on your hardware or inside a tenant you control, and earns its place once data sets need rebuilding routinely. Four questions settle which shape you need:
Pin it down: questions for your evaluation
-
Which Blockify deployment does our quote cover: the on-device utility, or a server sized for our volume?Which scale of the pipeline you can run in production.
-
If we build a data set with basic chunking instead of a full Blockify run, which accuracy figures still apply?Whether the numbers you were shown attach to the data set you plan to build.
-
In what format does a data set leave the pipeline, and which of our applications can consume it?Whether the ingestion investment outlives any single assistant.
-
Which model ships on the machines we deploy, and can we swap it without rebuilding data sets?That model choice and data preparation stay independent decisions.
Inside the Assistant: Switching a Data Set On and Off
The hand-off from stage three to stage four is a switch, and the switch is visible. A completed data set arrives through the add-data-set control, or gets built on the spot by running local files through the on-device Blockify utility. Once loaded it sits in the data sets list behind a toggle to the left of its name; blue means on. The same control sits beside the paper clip in the chat, so a data set drops out mid-thread without losing the conversation.
Load as many as the work needs. Iternal states there is no limit on how many data sets a user keeps or how many topics they cover, and they can be versioned — an IT data set beside an IT version-two data set. When a response draws on loaded material, AirgapAI marks it with a cube icon and shows the block pairings the model judged most relevant.
With everything switched off, the assistant says so. A chat with nothing active carries a pill reading exactly that, and clicking it turns one on. Ask in that state and the answer comes from the loaded model alone, which may know nothing recent and nothing about your organization. Users reporting that the assistant cannot find their own policies have usually left the toggle off.
What Each Component Is Not
Architecture confusion comes from one part being credited with another's job. Six corrections, one line each:
- Blockify is not a model trainer. It prepares data for models; training and fine-tuning sit outside it.
- Blockify is not a storage layer. It stays agnostic about where data is kept and retrieved; the output is text with governance metadata.
- Blockify is not a cure for missing sources. Iternal is explicit that it is no silver bullet.
- AirgapAI is not a language model. It is the application around one, built model agnostic so you bring your own.
- A data set is not a live connection. It is an artifact built at a point in time; when sources move on, somebody rebuilds it.
- AI Assist is not different software. It wraps AirgapAI in white-glove deployment, enterprise support and AI Academy training, as a bundle offered exclusively with Dell.
Read those six backwards and you have a purchasing test: a requirement landing outside all three parts needs a different component, not a different configuration.
Answered elsewhere
- What the pipeline accepts as input, and how content shape changes the result — see the input-formats page.
- How blocks are indexed and searched once loaded, and how large a corpus can grow — see the retrieval-architecture page.
- Grounding answers in your documents and showing which source produced each one — see the traceable-answers page.
- Automated crawling of file shares and object stores, and who owns re-indexing — see the corpus-currency page.
- Policy, permissions and federation when material sits across many systems — see the governance page.
- Plain-language definitions of the terms used across these pages — see the AI glossary.
FAQ
FAQ: How the Pieces Fit Together
Source documents go through the ingestion pipeline, which produces a data set; that data set is loaded into the local assistant; the assistant answers against the blocks you switched on, using a model running on the device or on a server you own. Iternal names those parts Blockify, the data set and AirgapAI.
Yes. Iternal builds Blockify architecture agnostic: it drops into an existing stack without changing the vector database, the re-ranker or the extraction step. The output is assistant agnostic too, so a processed data set can be consumed outside AirgapAI by any application.
It does, with a narrower result. With nothing switched on, the chat shows a pill saying no data set is active, and the answer comes from the loaded model alone. The model setting carries general knowledge and tone; your data set carries your facts.
Import a completed data set through the add-data-set control, or build one by dragging local files into the lightweight Blockify utility that ships with AirgapAI. Loaded data sets appear in a list with a toggle to the left of each name; blue means on.
AI Assist is a package rather than a separate application. Iternal and Dell built it as a Dell-exclusive bundle: AirgapAI software, white-glove deployment consulting, AI Academy training and dedicated enterprise support. The software in the diagram stays the same.
A data set is required, and the full Blockify run is the recommended way to make one. AirgapAI also supports basic chunking as a faster alternative, and a data set built that way is still stored as blocks. Ask which accuracy figures apply to the method you choose.
Draw Your Own Version of the Diagram
Take the five stages, write your own document sources into stage one, and name the owner of every arrow. The exercise surfaces what architecture questions are really about: who rebuilds the data set when sources move on. Settle that, and the rest is configuration.