What packs can do today, and what they cannot

The same loop — open the pack, read the map once, then ask — on these prepared packs:

• Industrial assets — plant-1, a synthetic refinery unit (about 258,000 facts)
• Venture funding — public venture and funding records (about 21 million facts)
• Corporate entities — public legal-entity records (about 86 million facts)
• Geographic knowledge — country records from the CIA World Factbook (about 111,000 facts)
• Pharma — public drug-identity records (about 6 million facts)

That shows one Knowledge Pack structure can represent very different domains and be queried the same way — once the knowledge is already in a pack. It does not show we can build that pack from messy production systems.

We will demonstrate reliable packing of messy, multi-system enterprise data. Packs do not update themselves as source systems change. We can then publish a head-to-head comparison against retrieval and generated queries. These are precisely the questions the four-week evaluation is designed to produce on a customer's own data.

If the source is already fairly clean — a consistent extract, a filing, a register, a well-structured table — an agent can help convert it. A person still has to own the model: what is a location versus a piece of equipment, which identifier locks an entity.

That is not the same as pointing an agent at twenty years of ugly data from dozens of systems and getting a trustworthy pack back. Getting messy, conflicting enterprise systems into a pack is the next thing to test, not a feature we ship today.

The pack can keep both. System A saying “installed in 2018” and System B saying “installed in 2019” do not have to overwrite each other. Each can keep its source, its dates, and the rest of its context.

The agent can then show the conflict and the basis for a choice. Automatically picking a winner is not something we have demonstrated. An evaluation should include at least one known conflict and an agreed rule for how the pack and the agent should behave.

Yes. Attribution that the agent can follow — this installation attested by this notification, this parts revision by this change number — is in the packs we have run. Numeric confidence and importance can sit on facts too. They are not a substitute for a business rule about which source wins.

No. It can hold both facts and the context on them. Who is authoritative is a modelling and policy choice for that domain, not an engine we claim to ship.

Each Knowledge Pack is a versioned, immutable snapshot. A pack is a hashed extract. The same question against the same revision should give the same answer. A new revision is produced when source data changes. That is deliberate: same question, same pack, same answer.

Keeping a pack continuously in sync is not the current product.

A live layer over a warehouse or lakehouse is a possible later architecture, not something we sell today. If it were built, the history of live changes would sit in that platform's audit trail, not inside a versioned pack. That would be a different product, not a leftover feature.

It makes unsupported answers harder: the agent asks for a named record from a bounded pack, and if nothing is there it gets an empty result, not invented prose. We have not published an independent benchmark of hallucination rates, or a controlled comparison with modern retrieval. Smaller answers are likely when you return a few hundred tokens instead of a dump of search hits; that should be measured on the partner's twenty questions. The scorecard is how we find out.

• Retrieval (RAG) finds similar text. A pack locks an identity and returns a record. Retrieval is still useful for finding which pack to open, or for reading narrative that was never structured.
• A graph database is a live store with a query language for analysts and applications. A pack is a file you hash and attach. If you already use Cypher, you keep it.
• A live connection (API) to SAP or another system of record gives current data and potentially too much authority. A pack is an extract, reproducible against a named revision, with no standing credentials in the agent's hands.
• Generated SQL asks a model to write queries. That is wrong too often to rely on for business answers.

Access isn't knowledge. System access lets the agent reach source data. It still has to work out what the data means, how it connects, what was true when, and where it was sourced.

• App connectors — Gmail or Drive read the user's own world. They don't hold reference knowledge.
• API wrappers — Add a tool for every endpoint, and choosing the right tool is where agents slip.
• Search — Returns relevant passages, with no guarantee that nothing is missing.
• Databases — Have full query power, but the schema doesn't explain itself.

Knowledge Packs work differently. Every pack uses the same small set of tools, so adding new knowledge doesn't add new tools. The pack describes itself: what it holds, how records link, and starter queries that show how and where to find records. Records come back structured, with time and source on every link. If anything was cut, the answer says so.

Packs complement Gmail and Drive; they don't replace them.

No. It means the pack does not contain that relationship. Public legal-entity records are the clear example: a name search can return more branded organisations than have an explicit management link in the pack. The right behaviour is to say what is represented, what is not, and which questions need another source — not to invent the missing link.

No. A versioned snapshot is not a confidentiality product. “The data stays with the client” means the default evaluation runs on their machine, without a First Cognition cloud ingest. That is not HIPAA, encryption, or zero-knowledge. Attribution on a fact is not a signed audit certification.

Prove the difference on one client's own data. A four-week evaluation against 20 real questions, measured against their current approach: one bounded domain and a masked extract on their machine. This is a paid evaluation. We run a small number of these at a time. We are not collecting data through this site, and we are not claiming the enterprise problem is already solved.

Book a 20-minute demonstration. We use synthetic refinery data on that call. Do not send production extracts. If it looks like a fit, we agree evaluation terms before any client data is used.

Path: 20-minute demo → four-week evaluation → report scorecard.