Skip to main content
← Projects / Case 02 Working ingestion + retrieval

Legal · citable retrieval

NexusLaw

Legal answers are only useful if you can cite them. NexusLaw is retrieval over primary law built around that fact — every passage it returns can name the exact statute, rule, or case section it came from — with an attorney intake pipeline alongside it.

Scroll through the system

01 / Origin

The challenge

Make legal research answerable and quotable. A general-purpose chatbot that summarizes law without a citation is worse than useless in practice — it cannot be relied on or checked. The goal was retrieval where every result is anchored to its source, over a corpus that is local and always available rather than dependent on slow external services at query time.

02 / Artifacts

What exists because of it

  1. 01

    Ingestion of Illinois primary law — statutes, court rules, and caselaw — plus practitioner reference material, into a searchable knowledge base.

  2. 02

    Section-anchored chunking: passages are split along sections, chapters, and numbered headings, carrying the metadata that makes a result quotable rather than sliced across two provisions.

  3. 03

    A drop-in ingestion path — add a new reference PDF and it is indexed automatically.

  4. 04

    An attorney intake pipeline in a CRM for the matter/relationship side of legal work.

03 / System

How it moves

Sources are ingested once and embedded into a vector database, so lookups do not gamble on a slow or rate-limited external API at query time. The load-bearing choice is section-aware chunking: legal text is addressed by citation, so chunks are anchored to sections and carry book/chapter/section metadata — that metadata is what makes a retrieved passage citable. Retrieval is exposed as a tool the assistant queries, rather than hitting the database directly, and a separate CRM handles attorney intake.

  1. Input Primary law statutes, rules, caselaw
  2. Step Section-aware chunking carries citation metadata
  3. Step Vector search Qdrant + local embeddings
  4. Output Cited answer names its section
Primary law is ingested with section-aware chunking, embedded for search, and every answer names the section it came from.

04 / Decisions

The tools and their jobs

  • Qdrant vector database
  • Ollama local embeddings (mxbai-embed-large)
  • Section-aware ingestion citable chunking + metadata
  • MCP tool retrieval interface
  • Twenty attorney intake pipeline (CRM)

05 / Today

An honest current state.

Ingestion and citable retrieval work across the Illinois primary-law sources and reference material, and new material can be added by dropping it in. The design principle — never trade section-anchored citability for convenience — has held. It is engineered as a working system rather than a documented product; the retrieval quality is the point, and it is real.