Product vision

Factflow is a domain-agnostic AI pipeline orchestration platform: it turns streams of heterogeneous content — web pages, office documents, chatbot exports — into queryable knowledge, reliably and with complete provenance. Every capability is a pluggable adapter wired by configuration, carried over a message broker, with storage, lineage, and replay underneath.

It is deliberately not a knowledge product itself. Products like DNB Kunnskap are built on Factflow. Factflow knows about adapters and messages; it knows nothing about knowledge — and that line is the heart of this page.

Why it exists

Enterprise knowledge in its native state is hyper-fragmented, hard to reach, and quietly self-contradictory — scattered across silos, locked behind interfaces that return documents instead of answers, inconsistent when two sources disagree. Turning that into something queryable, reliable, and trustworthy is hard, and one-off scripts don't get you there: they can't persist work, scale out, replay, or prove where an answer came from.

So Factflow was built as general machinery, not another one-off. The capabilities that matter — discovering and fetching sources, converting and segmenting them, embedding and retrieving them — are worth far more on a dependable, swappable substrate than welded into a single app. The hard engineering (a provider-agnostic broker, a reactive engine, durable replayable storage, end-to-end lineage) is the platform; the knowledge work lives at the edges as workflows. Because that knowledge methodology is configuration over the engine, a real product like DNB Kunnskap can be built on Factflow without rewriting it.

What it does

The canonical flow Factflow runs — most pipelines use only the slice they need:

Discover sources → Fetch → Store (raw) → Convert to markdown → Segment
→ Embed → (optionally) classify / synthesize → publish / serve

See The pipeline spine for how a message actually moves through it.

The value

Five things ad-hoc scripts and single-purpose tools can't give you:

• Unstructured → queryable. Messy, heterogeneous sources become a structured, searchable, embeddable knowledge substrate.
• Trust through provenance. Every output traces to its source via lineage and config snapshots — a claim that can't be traced is a claim no one will trust.
• Reproducibility & recovery. Storage is an event log; runs are replayable, so the same input deterministically yields the same output and partial runs can be resumed.
• Governance in the flow. Sensitive content is redacted, classified, and gated before it reaches a model — see Hygiene.
• Vendor & scale freedom. Provider-agnostic brokers, storage, and models; horizontal scale via competing consumers.

Platform vs product: Factflow inside DNB Kunnskap

The distinction people most often blur: Factflow is the engine; DNB Kunnskap is a product built on it. This diagram shows the nesting — sources flow in on the left, the FactFlow engine ingests/converts/cleans/detects/consolidates them, and Knowledge, Persons, and Concepts flow out on the right to downstream consumers.

The whitepaper names four roles, which map directly onto the picture:

Role	Name	What it is (in the diagram)
Engine	Factflow	the inner box — pluggable adapters, broker, storage, lineage, replay. "Knows nothing about knowledge; it knows about adapters and messages."
Workflow	Scribe	the knowledge methodology as configuration over the engine (the stage wiring + a taxonomy)
Artifact	Avalon	the knowledge base — the outputs on the right (Knowledge, Persons, Concepts); its own repository
Consumer	downstream	the bottom row — people, systems, and agents that use the knowledge

DNB Kunnskap is the whole outer box: Scribe + Avalon + the vision. The boundary is enforced in the architecture itself — the platform never depends on the workflows built on it; the knowledge workflows import shared platform services, never the reverse.

Belongs to Factflow (platform)	Belongs to DNB Kunnskap (product)
Engine, adapters, routing, broker, storage, lineage, replay, executions	The Scribe & Avalon pipeline configs and the taxonomy
Generic adapters: crawl, fetch, html→markdown, segment, embed	Knowledge adapters: concept detection, consolidation, knowledge-diff
The hygiene mechanism (an adapter slot)	The hygiene policy: sensitivity tiers, the org-wide-knowledge invariant, source precedence
Provenance as a property of transport	The Avalon corpus, its conflicts, its Persons & meta-knowledge views

The one-sentence test for anything new: if it's about moving, transforming, tracking, or replaying messages, it's Factflow; if it's about what the knowledge means — concepts, taxonomy, sensitivity policy, the corpus — it's the product on top.

Where to go next

• Overview — the technical mental model and package tiers
• The pipeline spine — how a message flows end to end
• Adapters & routes — the unit of work and how it composes
• Knowledge & Avalon synthesis — the product workflow in practice