Legal AI copilot architecture: retrieval, graphs, and an evidence path

Three different problems

Finding a similar passage, reconstructing relationships, and justifying a decision are not the same task. Vector retrieval ranks candidates by similarity. A graph can record explicit links between documents, events, parties, and rules. An orchestration layer routes work and records execution. None of those layers independently establishes legal correctness.

This edition updates an earlier Medium review. Tool names in the original article are retained only as historical context; they are not a current recommendation and do not establish any tool’s suitability without a separate assessment.

Retrieval as a candidate list

A similarity result should begin a control process rather than end it. The index must expose the corpus version, jurisdiction, effective date, and document identifier. Deterministic filters can reject out-of-scope material before a passage reaches a model. A result without those fields may sound relevant while referring to the wrong authority or time period.

Graphs as explicit structure

A graph becomes useful when relationships carry domain meaning: a document concerns a matter, one event precedes another, or a rule applies during a defined period. An edge should point to a source or rule explaining why it exists. A graph without provenance merely moves uncertainty into a more persuasive visual representation.

[corpus + version]
       ↓
[retrieval + filters] → [candidates]
       ↓
[event and rule graph]
       ↓
[citation-grounded synthesis]
       ↓
[test + professional decision]

Orchestration and accountability

An orchestrator should not hide tool boundaries. It records parameters, the selected path, failures, and the hand-off to a person. For a high-risk task, a safe result may be to stop and identify missing material instead of generating a complete-looking answer.

The evaluation set

A scenario defines the user question, admissible sources, expected citations, relationships to reconstruct, and an unacceptable failure. Retrieval metrics remain separate from citation fidelity, completeness, and abstention decisions. One average score conceals where the workflow failed.

A minimal recommendation

Start with one workflow and a source register. Add simple filtered retrieval, a citation check, and an execution trace. A graph or another agent is justified only when it addresses a named failure that can be measured again. The resulting system supports a professional; it does not become a source of legal advice.