01
Capture
Exact URLs, pages, source IDs, text, JSON-LD, images, CAD, and source revisions.
Resource · Import architecture
The importer is an evidence pipeline, not a scraper.
Every import path should preserve where data came from, how confident the mapping is, what was inferred, which target identity it will affect, and which human approved the final Apply.
02
Normalize
Convert source-specific evidence into product candidates, Product Types, fields, relationships, and identity matches.
03
Apply safely
Preview and human review precede idempotent revision writes to the active product system.
AI vs deterministic
Choose interpretation only where interpretation is necessary.
ConfigAI should not route every source through an LLM. Structured providers and deterministic APIs remain direct; research agents focus on sources intended for people.
Deterministic
pCon Catalog
Direct provider product list, hierarchy, metadata, and CAD conversion where available.
Structured
pCon / OFML
Provider-authorised supplier structures and configuration data through dedicated source adapters.
Agent-assisted
Website
Bounded JEV capture plus agent reasoning for page discovery, evidence, product identity, and relationships.
Agent-assisted
PDF / conversation
Extract evidence first, then let the Product Setup workflow decide what requires clarification.
Replay & cost
Research once, replay evidence many times.
JEV-backed capture can reduce repeated model context by storing bounded evidence results and replaying them when Product Setup or later correction work needs the same source.
01Bound the requestExact link priority, page limits, timeout, and evidence classes constrain research cost.
02Parallel captureFetch independent pages together when it improves latency without changing the durable import task.
03Persist evidenceKeep captured text, structured metadata, images, low-confidence findings, and deferred links.
04ReplayA repeated import or correction request reuses the evidence before calling the provider again.Cost control
Identity & revision
Every proposal must know whether it creates, reuses, or revises.
Import quality is not just field extraction. It is matching source identities to durable ConfigAI products and making the semantic impact visible before mutation.
Source identity discovered
Provider ID, exact URL, article number, source family/series, or other stable evidence is captured.
Target match proposed
Create, exact reuse, or source update is computed against existing ConfigAI identity and product revisions.
Semantic impact previewed
Field/Product Type changes are shown separately from simple value changes.
Review
Idempotent Apply
The approved mutation writes one durable outcome and replay does not duplicate products.
Stable
ConfigAI
Import once. Keep the source relationship for every future revision.
A reliable product system remembers where the data came from and how it became the reviewed ConfigAI definition.