Citation-Grounded RAG for Legal Research and E-Discovery

The compliance context: privilege, sanctions, and the duty of competence

Legal is governed less by a data-protection statute than by professional-responsibility rules that carry personal consequences for the lawyer. ABA Model Rule 1.1 and its state analogues impose a duty of technological competence, and Model Rules 5.1 and 5.3 make partners responsible for supervising non-lawyer assistance, which now includes AI output. Rule 11 and its state equivalents expose the signing attorney to sanctions for citations that do not exist or do not stand for what they are cited for. Any system we build has to assume its output may end up in a filing and must make verification the path of least resistance, not an optional extra step.

Confidentiality raises the stakes further. Model Rule 1.6 and attorney-client privilege mean client matter data cannot be used to train a shared model or leak across matter boundaries, and inadvertent disclosure during e-discovery can waive privilege under Federal Rule of Evidence 502. We build with strict matter-level isolation, no training on client content by default, and audit logging of every retrieval so a firm can demonstrate what the system accessed and when. Where a matter touches EU personal data or health records, GDPR and sector rules layer on top, and our SOC 2 aligned controls and GDPR-ready data handling are designed to satisfy an outside general counsel's due diligence, not just an engineer's.

Why RAG, not fine-tuning, and where an agent earns its place

Fine-tuning teaches a model a style or a vocabulary. It does not give the model a reliable, updatable memory of which cases are still good law, and it makes the source of any given claim impossible to point at. For legal research that is disqualifying. Precedent shifts, statutes are amended, and an opinion can be overruled or distinguished next term, so the authoritative content has to live in a retrieval index that a firm can update the day a decision comes down, not baked into weights that are months stale. RAG is the correct spine because it separates the reasoning layer from the authority layer and lets us cite the authority directly.

The honest distinction is retrieval-and-memo versus agentic workflow. A memo drafting task, summarize the holding across these twelve authorities on a narrow question, is a bounded retrieval problem that a well-tuned RAG pipeline handles with high reliability and low latency. E-discovery triage across a million documents is closer to an agentic workflow: iterative querying, relevance and privilege classification, and routing, where the system plans multiple retrieval passes and hands borderline calls to a reviewer. We use agents where the task genuinely requires multi-step planning and tool use, and we resist them where a single grounded retrieval is more predictable and easier to audit. Choosing that boundary deliberately, rather than making everything an agent, is the difference between a system a firm trusts and one it quietly stops using.

Named use cases and honest outcome ranges

Precedent and authority search: an assistant that answers a research question with a ranked set of on-point authorities, each with the quoted passage and pincite, and flags negative treatment where the firm's citator data supports it. First-pass research time on a defined question commonly compresses by 40 to 70 percent, with the review step preserved rather than removed, because the lawyer verifies quotes rather than hunting for them.

Memo and brief-section drafting: grounded first drafts that assemble the retrieved authorities into an argument skeleton with every proposition linked to a source. Drafting throughput on routine research memos improves in the range of 30 to 60 percent, though we scope this conservatively because the value is in a verifiable draft, not an unread one.

E-discovery triage and privilege review: classification of production documents for responsiveness and privilege, prioritizing a review queue so human reviewers see the likely-relevant and likely-privileged material first. Reviewer efficiency gains in the 45 to 90 percent range are consistent with what technology-assisted review has shown for years, and RAG adds explainability by attaching the passages that drove each call. We present these as ranges because they depend on corpus quality, matter type, and the review protocol, and we measure the real number per engagement rather than promising a headline figure.

Integration realities: what it connects to and how the data flows

A legal RAG system is only as good as the corpus it retrieves from, so integration is most of the work. On the authority side that means the firm's licensed research content and internal brief bank, and where terms permit, connectors to the platforms a firm already pays for. On the matter side it means the document management system, commonly iManage or NetDocuments, and the e-discovery platform, commonly Relativity, where productions and coding decisions already live. We ingest with document-level and matter-level access controls preserved end to end, so the assistant can never retrieve across an ethical wall it should not cross.

The pipeline is deliberate: parse and OCR mixed formats including scanned exhibits and email families, chunk with structure awareness so a holding is not split from its reasoning, embed, and index with metadata for jurisdiction, court, date, and matter. Latency matters because a lawyer researching interactively will not wait, so we engineer retrieval and generation to return grounded answers in seconds, not tens of seconds. The evaluation loop is not optional: we build a curated set of research questions with known-correct authorities, and every change to prompts, chunking, or models is scored against it before it ships. Evaluation by default is how we keep a system honest as the corpus and the models underneath it change.

Failure modes and guardrails: making hallucination structurally hard

The failure mode that ends careers is the confident fabricated citation, so we design the system so it cannot cite what it did not retrieve. Answers are constructed from retrieved passages, every claim carries a source link back to the exact document and pincite, and the interface makes the quoted text visible next to the assertion so verification is a glance, not a search. When retrieval returns nothing on point, the correct behavior is to say so and hand off, not to synthesize a plausible-sounding answer, and we tune and test for that refusal explicitly.

Guardrails run at several layers: a citation verifier that checks every referenced authority resolves to a real document in the index and rejects any that do not, jurisdiction and date filters so a superseded statute or an out-of-forum case is flagged, and privilege classifiers with conservative thresholds so the cost of a false negative in e-discovery is borne by extra human review rather than by waiver. Human oversight is a permanent part of the design, not a training-wheels phase, because the professional-responsibility rules put the lawyer on the hook regardless of the tool. We take end-to-end ownership of this stack, from retrieval to eval to the audit trail governance needs, so a firm has one accountable partner for a system whose mistakes have real legal weight.

Common questions

Can the assistant guarantee it will never cite a case that does not exist?

It is designed so that citations are drawn only from documents actually retrieved from the firm's indexed corpus, and a verification layer rejects any referenced authority that does not resolve to a real document in the index. That structurally prevents the fabricated-citation failure that has drawn Rule 11 sanctions with general-purpose models. It does not replace the attorney's duty to verify, and we build the interface so that verification takes seconds rather than removing it.

How do you keep one client's privileged material from leaking into another matter?

Matter-level isolation is enforced at retrieval time. The index carries access-control metadata mirrored from the document management system, so the assistant can only retrieve documents the querying user is entitled to see, and ethical walls are respected end to end. Client content is not used to train shared models by default, and every retrieval is logged so the firm can demonstrate exactly what the system accessed, which matters for both Rule 1.6 confidentiality and FRE 502 waiver risk.

Is this a replacement for our e-discovery review platform like Relativity?

No. It integrates with the platform you already use rather than replacing it. The RAG layer adds triage and explainable classification on top of your existing productions and coding decisions, prioritizing the review queue and attaching the passages that drove each responsiveness or privilege call. Reviewers still make the final determinations, and the coding decisions flow back into your platform of record.