Contract review and clause extraction

Architecture and data flow

Organisation size

10–500 employees

Data classes

confidential, privileged, personal

Constraints

contracts are privileged or commercially sensitive and must stay in-house; a defined clause set worth extracting — not "review everything"; a lawyer who will check the output and own the clause definitions; a 48 GB GPU for a 32B model; 24 GB works for a narrower clause set at lower accuracy

Industries

Legal, Professional services, Financial services, Real estate

Jurisdictions

any

• On-premise single 24 GB GPU server

  GPU

  NVIDIA RTX 4090 24 GB (or NVIDIA L4 24 GB for a rack-mounted, 72 W alternative)

  VRAM

  24 GB

  System RAM

  64 GB

  Storage

  2000 GB

  CPU

  16-core x86 server CPU (AMD EPYC 7003/9004 or Intel Xeon Scalable)

  Form factor

  Tower server

  Indicative costUS$4,000 – US$9,000

  indicative build cost for the complete machine, USD, Aug 2026 — verify with a local supplier

  The default box for a 20–60 person firm. Fits a 14B model at 4-bit with roughly 8 GB of KV cache left for concurrent chat, or an 8B model at fp16. NVMe storage sized for the model cache plus a document corpus and its embeddings. Add a UPS and an offsite backup target — this machine holds the whole knowledge base.

• On-premise single 48 GB GPU server

  GPU

  NVIDIA L40S 48 GB (rack) or NVIDIA RTX 6000 Ada 48 GB (office workstation)

  VRAM

  48 GB

  System RAM

  128 GB

  Storage

  4000 GB

  CPU

  24–32-core x86 server CPU (AMD EPYC or Intel Xeon Scalable)

  Form factor

  Rack server

  Indicative costUS$12,000 – US$22,000

  indicative build cost for the complete machine, USD, Aug 2026 — verify with a local supplier

  The step up when answer quality matters more than price: a 32B model at 4-bit with generous KV cache, or a 14B model at fp16 with a long context. ECC memory and a standard 2U chassis mean it belongs in a rack or comms room rather than under a desk.

Dockerdocker

infrastructure

Linuxlinux

operations

LLM evaluationllm-evaluation

ml

LLM inferencellm-inference

ml

OCR and document parsingocr

ml

PostgreSQLpostgresql

data

Prompt engineeringprompt-engineering

ml

Pythonpython

development

Retrieval-augmented generationrag

ml

Security hardeningsecurity-hardening

security

Commands are copied from each project’s own documentation, and the page they came from is linked under the step. 5 of 7 steps currently open an evidence record. The rest are linked to their source; §10 says which of those documents were fetched and which were fetched without their anchor being found — two different states, named differently there.

1. 01

   Define the clause set with a lawyer, on paper, first[2]

   Assessment

   List the fields: parties, term, termination, liability cap, indemnities, governing law, assignment, confidentiality, data protection. For each, write what a correct extraction looks like and what an ambiguous one looks like. This document is the specification, the prompt and the test set — no code should be written before it exists.

2. 02

   Convert contracts with Docling[4][3]version-sensitive

   Assessment

   Install Docling and convert a sample of real contracts, including the scanned ones. Inspect the markdown output: if the tables and clause numbering survive, extraction will work; if they do not, fix conversion before touching the model.

   pip install docling

   docling https://arxiv.org/pdf/2206.01062

   from the Docling README

   Source documentation

3. 03

   Serve the model with vLLM[1]version-sensitive

   Assessment

   Start the official image with the GPU attached, sized for a 32B model at 4-bit on a 48 GB card. Set an API key. Keep the context length high enough for the longest contract section you will pass, not the longest contract — chunk by clause, not by document.

   docker run --runtime nvidia --gpus all \
       -v ~/.cache/huggingface:/root/.cache/huggingface \
       --env "HF_TOKEN=$HF_TOKEN" \
       -p 8000:8000 \
       --ipc=host \
       vllm/vllm-openai:latest \
       --model Qwen/Qwen3-0.6B

   from the vLLM docs — substitute the model you sized for

   Source documentation

4. 04

   Extract against a JSON schema, not into prose[2]version-sensitive

   Assessment

   vLLM "supports the generation of structured outputs using xgrammar or guidance as backends". Send each clause section with the schema for that clause and require a source quote plus page number in the schema — a field the model cannot fill without pointing at the text is a field a reviewer can check in seconds.

   Source documentation

5. 05

   Store extractions and embeddings in PostgreSQL

   Assessment

   One row per clause with the contract id, the field, the value, the quote, the page and the model version. Embed the clause text so "have we agreed this liability cap before?" becomes a query rather than a memory test.

   CREATE EXTENSION vector;

   Source documentation

6. 06

   Put a human in the loop, visibly

   Assessment

   Every extraction starts as a draft with a reviewer, a decision and a timestamp. Show the source quote beside the field. Record corrections: the corrected set is both your accuracy measure and the evidence for whether the tool is earning its keep. A purpose-built review screen beats a chat box for this; Open WebUI, cited below, is the fast way to have something in front of a lawyer this week, and it is optional in the sense that a small internal app replaces it — not in the sense that the review step is.

   Source documentation

7. 07

   Measure per field, not per document[2]

   Assessment

   Take 50 contracts a lawyer has already marked up. Report precision and recall for each field separately. A pipeline that finds governing law perfectly and misses indemnities is not "85% accurate" — it is ready for one field and not the other, and only a per-field table shows that.

Structured issue-spotting to support your own review — not legal advice. Verify against the cited primary sources and your counsel.

• A specialist legal AI vendor

  Vendors in this space ship the clause library, the review UI and the evaluation work you would otherwise build. The trade is that privileged documents go to a processor.

  • — Far less to build; domain-tuned out of the box
  • — Privileged material leaves the firm — DPA, residency and client consent questions follow
  • — Per-seat or per-document pricing rather than a one-off build

  harvey

• Start with retrieval only

  If the real need is "find the contract with this clause", a knowledge base answers it in a fortnight without any extraction schema at all.

  • — Much simpler and quicker to value
  • — No structured clause table, so no comparison or reporting across contracts

  open-webuivllmpgvector

• Self-hostedPrivate company knowledge base (self-hosted RAG)Open WebUI as the employee interface, vLLM serving a Qwen2.5-14B-Instruct model on a single 24 GB GPU, PostgreSQL with pgvector for chats and embeddings, and OIDC single sign-on — all in Docker on one server in your office or colocation rack. Ollama replaces vLLM for teams under about 20 users; a 48 GB GPU lets you run a 32B model for better answers.

1. 01
   Tier 2·official vendor legal documentationVendor legal and technical documentation

   Using Docker - vLLM

   https://docs.vllm.ai/en/latest/deployment/docker/

   Fresh·Retrieved 25 Aug 2026sha256:67e857491614

   1 record01 vLLM offers an official Docker image…

2. 02
   Tier 2·official vendor legal documentationVendor legal and technical documentation

   Structured Outputs - vLLM

   https://docs.vllm.ai/en/latest/features/structured_outputs/

   Fresh·Retrieved 25 Aug 2026sha256:1d9d9841283c

   1 record02 vLLM supports the generation of structured…

3. 03
   Tier 2·official vendor legal documentationVendor legal and technical documentation

   Index - Docling

   https://docling-project.github.io/docling/usage/

   Fresh·Retrieved 25 Aug 2026sha256:b695cbd95509

   1 record03 from docling.document_converter import DocumentConverter

4. 04
   Tier 2·official vendor legal documentationVendor legal and technical documentation

   GitHub - docling-project/docling: Get your documents ready for gen AI

   https://github.com/docling-project/docling

   Fresh·Retrieved 25 Aug 2026sha256:7f9d5e4fa303

   1 record04 docling https://arxiv.org/pdf/2206.01062

Not yet in the ledger

1. vLLM — Quickstart (online serving)

   Docs

   not yet fetched

2. pgvector — README (installation and getting started)

   Readme

   not yet fetched

3. Open WebUI — Quick Start (Docker)

   Docs

   not yet fetched

4. Open WebUI — Hardening

   Docs

   not yet fetched

Deployed this stack, or hit something this page does not cover? Corrections, sources and implementation reports are what keep a recipe worth reading.

If you would rather not build it, we can introduce a forward-deployed engineer who has deployed this stack before. The enquiry form starts from this recipe.