A research assistant for Australian immigration law. Retrieval follows the citation chain through a 99,854-relationship knowledge graph, and Claude 3.7 Sonnet writes one cited answer.
Konze — Aussizz Group
Immigration & legal services, Australia
Claude 3.7 Sonnet
3.5 Sonnet · 3.5 Haiku selectable
Active — validated, hardening for production
Reckonsys paired Claude with a purpose-built legislative knowledge graph:
How it works
Semantic search finds a relevant passage. It does not know that the passage is qualified by a provision three citations away — the knowledge graph does.
So retrieval starts from the relevant passage and walks the citation chain outward, gathering the provisions a researcher would have followed by hand.
Australian immigration law is unusually hard to research. It spans four document types — Acts, Regulations, Policy and Legal Instruments — that are updated constantly and cross-reference each other heavily. Answering a client question correctly rarely means finding one passage; it means following the citations from that passage to the provisions that qualify it.
A knowledge graph of the legislation. 21,727 legislative documents and 99,854 citation relationships modelled in Neo4j, so the structure of the law is queryable rather than implicit in the text.
Semantic retrieval over the corpus. 114,162 embedded legislation chunks across Acts, Regulations and Legal Instruments, indexed for meaning rather than keyword match.
Graph traversal at query time. Retrieval starts from the semantically relevant passage, then walks the citation chain outward, gathering the provisions a researcher would have followed by hand.
Claude synthesises across documents. Claude 3.7 Sonnet reads the assembled chain and produces a single cited answer, rather than handing back four documents for the agent to reconcile.
Cohabitation is not a strict requirement, but any period apart must be accounted for and the relationship assessed against the four statutory factors.
The model
Claude 3.7 Sonnet handles both halves of the pipeline: interpreting the legal question to decide what to retrieve, and synthesising the retrieved chain into one answer with every citation intact. Claude 3.5 Sonnet and 3.5 Haiku are selectable alternatives.
Claude 3.7 Sonnet is the default model for both halves of the pipeline: query analysis, which interprets the legal question and determines what to retrieve, and response generation, which synthesises the retrieved chain into one answer with citations intact. Claude 3.5 Sonnet and Claude 3.5 Haiku are integrated as user-selectable alternatives, so the workload can be matched to the question.
The reason Claude suits this work is the same reason the knowledge graph is necessary. A legal answer has to stay inside the retrieved provisions, attribute every claim, and decline to go further than the source supports. Synthesising across four documents that qualify one another — while keeping each claim tied to the document that supports it — is reasoning under constraint, and it is precisely where a confident but unsourced answer would be most damaging.
“Synthesising across four documents that qualify one another is reasoning under constraint.”
Evaluation
The platform compares 11 models across 4 providers on Aussizz's own immigration questions, not on benchmarks. With the client free to switch models at will, Claude 3.7 Sonnet as the default is a verdict rather than a vendor preference.
Validation
Tested end to end against real client-provided question sets across six visa categories — Partner, Parent, Student, Employer-Sponsored and Tourist among them. Not synthetic evaluation questions: the ones Aussizz's agents actually field.
Roadmap
Reckonsys continues as Konze's engineering partner on the programme.
Phase 1 · complete
Core Claude-powered retrieval and knowledge-graph pipeline built and validated against real client test cases.
Phase 2 · current
Production hardening — dedicated hosting, cost monitoring, and a legislative data-refresh cadence to keep the corpus current.
Phase 3 · next
Full integration with the Aussizz client platform, and coverage extended into policy, where ~97,000 more chunks are prepared for indexing.
Claude 3.7 Sonnet · Claude 3.5 Sonnet · Claude 3.5 Haiku · Claude API · Neo4j knowledge graph · vector embedding and semantic retrieval · citation-chain graph traversal · multi-provider model comparison across 11 models
Phase 2 covers production infrastructure hardening — dedicated hosting, cost monitoring, and a legislative data-refresh cadence to keep the corpus current as the law changes. Phase 3 is full integration with the Aussizz client platform, replacing the demo environment's auth stub with their live login, and extending coverage into the policy domain where a further ~97,000 chunks are prepared for indexing.
Let's collaborate to turn your business challenges into AI-powered success stories.
Get Started
