DocsGPT gives legal teams and legal-tech products retrieval with page-level citations across millions of pages and multiple languages, with guardrails and lawyer review as the final step. It powers LexEU's EU-law research over 2.5M+ pages in 6 languages and, at Balt Alnoor, IP drafting that went from 6+ hours to under 5 minutes. Legal-tech companies use DocsGPT as the retrieval and agent layer behind their own product.
Last updated:
Five stages: a client submission or the corpus comes in as filings, PDFs or images; the document parser produces text, with OCR or a vision model where a scan needs it; the model researches across the corpus or drafts against a template; the output is a cited answer or a near-final draft; a lawyer reviews and approves before anything leaves the system.
Citations
The central risk in legal AI is a hallucinated citation, so generation must be grounded in retrieved passages and every citation must link to a real page.
Languages
Multilingual legal terminology defeats general-purpose embeddings; retrieval has to be tuned to the domain and to each language the practice works in.
Review
AI output in legal work is a near-final draft, not a decision: a lawyer's review is the final stage of the pipeline, by design.
Workflow 01
Hybrid retrieval (BM25 + vector) finds the passage; GraphRAG follows multi-hop questions across regulations and case law; every answer cites title, source and page, and clicking a citation opens the exact passage beside the chat, with the file and source it came from. The Research agent handles multi-step questions with budget controls.
Hybrid retrieval (BM25 + vector) · GraphRAG · citation reader (title, source, page) · Research agent
Workflow 02
The default embedding model is multilingual, so a German statute and an Italian judgment land in one index and answer questions asked in either language; another local or remote embedding model can be swapped in where a domain needs it. Chunking and query rephrasing are set per source, and scanned pages are read with OCR in the languages you set.
Multilingual embeddings by default · local or remote embedding models · per-source chunking and rephrasing · OCR languages
Built for the client · fine-tuned legal embeddings and a domain fine-tuned LLM, as in LexEU.
Workflow 03
A Workflow takes the intake (images, PDFs, filings), extracts text with OCR or a vision model, has an AI Agent node draft a structured description, renders the template as a versioned DOCX or PDF artifact, and stops at a human-approval step before release. This is the IP drafting pattern. For a one-off matter, attach the filings in chat and keep the ones worth reusing with Add to Knowledge.
Workflow builder · OCR / vision extraction · AI Agent node · Artifact tool (DOCX, PDF) · human approval · Add to Knowledge
Workflow 04
Anti-hallucination prompts constrain generation to retrieved passages; guardrails run on retrieved sources and on the final answer (grounding in sources, which flags claims the passages do not support, plus banned terms, link policy and LLM-judge policies); human-in-the-loop gates responses and tool calls; per-domain benchmarks assert that answers cite the expected sources before a change ships.
Anti-hallucination prompts · grounding check · source and answer guardrails · LLM-judge policies · human-in-the-loop · docsgpt-cli bench
Five controls, in the order they act on a question. None is absolute on its own; together they make a wrong citation detectable and reviewable.
Grounding
Answers are generated only from retrieved passages; anti-hallucination prompts refuse when the corpus does not contain the answer.
Page-linked citations
Every citation carries title, source and page number; clicking it opens the exact passage beside the chat, with the file and source it came from.
Source guardrails
Guardrails check retrieved sources and the final answer: grounding in sources flags claims the passages do not support, alongside banned terms, link policy and LLM-judge policies.
flag · redact · block
Human-in-the-loop
Responses and tool calls can require approval; drafting ends at a lawyer's approval step.
Benchmark assertions on sources
A per-domain suite asserts that answers cite the expected sources before any change ships.
docsgpt-cli bench
application-draft.docx
artifact · append-only versions
Results from named deployments, with the source.
2.5M+ pages · 6 languages
Multilingual EU-law research over 2.5M+ pages, answers in 6 EU languages with full source citations, 100–200 ms retrieval.
Source: LexEU
6+ h → under 5 min
IP application drafting: 6+ hours to under 5 minutes per application, with lawyer review as the final step.
Source: Balt Alnoor
Managed is the usual posture here: a private instance without operating it.
Law firms typically take a dedicated managed instance or the Cloud plan — the IP drafting deployment runs on cloud models. Firms with residency rules run DocsGPT on-premises with a local model. Legal-tech companies run it in their own cloud as the backend of their product.
Scope, measure, then decide.
Select one corpus and one drafting type
One jurisdiction or practice area for research; one document type for drafting.
Lawyers write 30–50 benchmark questions
Each with the expected sources, so citation accuracy is testable from day one.
Ingest, configure retrieval, run the benchmark
Iterate on embeddings, chunking and rephrasing until the source assertions pass.
docsgpt-cli bench
Build the drafting workflow with a review step
Intake, extraction, draft, versioned artifact, lawyer approval.
Pilot with a small group
Measure citation accuracy and time per draft against the current process.
For legal-tech products the commercial model is a subscription plus an engineering retainer — see For AI product builders. Firms: Cloud at $20/seat/month or a Managed instance from $2,000/month — see pricing.
A 30-minute demo on your use case, or a free on-premises review with an engineer.