How Konneal works
The assets are the artifacts your organization has already authored. The passes are the pipeline stages that bind that structure into the indexes the service queries. The gates are the measurements that decide whether a change ships. The pipeline runs in one direction, and every stage leaves a record.
What you bring
Each asset below is structure the organization has already authored. Every one of them raises what the serving layer can do, and the library alone is enough to begin.
Your authored library holds documents written with clause structure, with typed tables, formulas and figures.
Your publication identifiers, parsed natively so questions naming a publication steer retrieval.
Your bibliographic records carry publication families, editions and supersession links.
Your terminology collection holds defined concepts, registers and multilingual spellings.
Your structured models are served as first-class objects.
Your machine-readable models of the rules carry constraints, calculations and test sequences.
The suite makes the ladder possible: documents authored in Metanorma give the corpus its spine; rules modelled in Primmel give it machines to run. Each layer of structure makes a specific range of capability possible, and the serving layer multiplies what is there.The measurement is documented on the annealment page.
The annealment passes
From publications to the index, each pass adds verified structure and leaves a record it can be replayed from.
01
Corpus assembly, with provenance
The corpus is assembled from the organization’s authored sources. Every chunk records where it came from, together with its publication, edition, language and a quality tier, so every later stage can be held to the record.
02
Clause-boundary chunking
Documents are split along their own clause structure — the structure the authors wrote — rather than by fixed-length token windows. Every chunk carries its clause number, so a citation can name an exact location in the source document.
03
Contextual enrichment
A one-time pass writes each chunk a short preamble stating where it sits: publication, part, clause, subject. The pass is recorded, so it replays exactly after any rebuild — enrichment lives in the index, not in transient artifacts.
04
Typed units
Tables, formulas and figures are extracted as first-class retrievable objects alongside the prose. A formula keeps its source notation; a table keeps its columns and rows as data; a figure keeps its original drawing.
05
Terminology and bibliographic binding
Defined terms link to the clauses that define them, and the bibliographic registry organizes publication families with their editions and supersession links, so which edition is in force is answered from the record. The registry is joined by a citation graph: what each edition’s bibliography cites is extracted as a graph edge, so a question of the form “which standards does this publication reference” is answered from the record rather than by hoping a text search finds the bibliography section.
06
The model plane
Where the publications carry machine-readable models, the models themselves are indexed: typed attributes, constraints with numeric limits, calculations and test sequences. This is what allows the system to execute a rule instead of quoting it.
Everything meets in one vector index under a declared contract: a single validated boundary enforces the wire schema and the legal metadata shapes, and public and access-restricted content live in physically separate indexes — separation that holds structurally, not by a filter that could be forgotten.
What happens when you ask
Seven stages, in order.
01
Understanding
A language model reads the question first: its language, any publication or edition it names, whether it asks about a defined term, and what earlier turns contribute. The same model makes these judgements for every question.
02
Retrieval
The question expands into several search forms — reworded variants and a draft answer as an additional key — run against a hybrid dense-plus-lexical index, fused by rank. Two structure-aware lanes add their own candidates: the terminology graph contributes the clauses that define a term when the question is about one, and the citation graph supplies the list of cited standards as structured data when the question asks what a publication cites, with the bibliography section retrieved as the passage that grounds it.
03
Ranking over structure
A cross-encoder orders the candidates. Because the corpus is a clause tree, a clause’s score also blends its neighbours’: a section whose clauses are collectively relevant rises. Evidence reaches the answer model in document reading order, and near-duplicates from one clause chain collapse.
04
Vocabulary binding
Everyday words rarely match defined terms. The terminology index links the member’s phrasing to candidate terms; the answer then leads with the corpus’s own term, quotes its definition and cites where it is defined.
05
Answering under contract
The answer is generated from the retrieved passages only, under a contract that requires an inline citation on every claim and verbatim quotes for normative values. Typed objects render from source; the model never retypes them.
06
Checking
A judge grades whether the passages support the answer, and a second check scores faithfulness. Failed checks trigger one corrected retry. Answers that remain unsupported are withheld rather than shown.
07
Conformance by execution
A question that names a machine-checkable rule gets the rule evaluated. Values bind to the rule’s parameters, the check runs deterministically, and the verdict — with the arithmetic shown — is attached as data the model must present faithfully.
What decides whether anything ships
A golden suite decides whether a change ships.
A golden question set with expected citations and witnesses runs end to end — three times — before any change is promoted, with paraphrase probes and leakage probes alongside. Grading is deterministic wherever possible: expected spans must appear in retrieved passages, and judged components are labeled as judged. The evaluation harness and the live service share a single judge implementation, so the measurement and the served behaviour cannot drift apart. A judge that cannot complete a judgement returns no score rather than a truncated one. A degenerate all-zero judgement is retried once rather than trusted.
The gates never run while the index is being modified, because an honest measurement requires a settled system. Every retrieval change bumps the index version, so no answer is served from a superseded index.
Deploys run from CI on merge, and a smoke probe questions the live service before a deploy is called done.
Two capabilities go beyond answering and can be tried directly: conformance by execution, and the checking of any answer against the corpus. Both run as in-page demonstrations onthe use cases page.
Platform and models
All models are open-weight. No corpus data is sent to third-party services, and the system has no dependency on a proprietary model provider. Different roles use different models, chosen for the job.
The reference deployment runs on Cloudflare — Workers, Vectorize, D1, KV, R2 — with reference adapters in the engine. The design runs anywhere that provides model serving, vector search and durable storage.
The service also exposes a conversational HTTP API and an MCP server for AI agents, with tools for grounded answers, passage search and the publication registry.
The narrative for the reference corpus, written for its readers, is atOIML SMART AI — how it works ↗.