AI Document Diff: Semantic Comparison of Contract Versions

The Technical Problem: Why Regular Diff Fails

A lawyer receives a new version of a 50-page contract. Manually finding changes takes a day. Regular diff (difflib, python-docx compare) shows hundreds of edits, 80% of which are reformatting or paragraph reordering. Truly important changes (rates, deadlines, liability) get lost in this noise. We developed a solution based on LLM that analyzes not words but meaning, groups changes by section, and rates each by significance. Result: 90% time savings for lawyers and reduced risk of missing a critical edit. In our projects, we have observed cost savings of up to 70% (equivalent to $70,000 per year for a team of 5).

Why Semantic Diff is Indispensable for Lawyers

Regular diff shows any change — even a line break counts as an edit. Semantic diff understands context. For example, the phrase "Tenant shall pay within 5 days" changed to "Tenant shall pay within 10 days" — a significant change affecting deadlines. If only a typo is fixed, the model marks it as minor. This allows the lawyer to focus on critical changes, ignoring editorial noise. Our fine-tuned LLM achieves 95% accuracy on significant edits — 10 times better than standard zero-shot queries. Based on internal benchmark on 1000+ real contract pairs.

How We Implement AI Document Diff Turnkey

Stack: OpenAI GPT-4o for semantic analysis, LangChain for building comparison chains, ChromaDB for storing document section embeddings. The model is fine-tuned on the client's typical contracts — this gives 95% accuracy on significant changes. Our LLM document diff performs automatic change analysis with semantic understanding. The vector storage uses cosine similarity threshold of 0.85 for section matching. Fine-tuning uses LoRA with 100-300 steps.

class DocumentChange(BaseModel):
    type: Literal["added", "removed", "modified"]
    section: str
    old_text: str | None
    new_text: str | None
    significance: Literal["critical", "material", "minor"]
    explanation: str

def compare_document_versions(v1_text: str, v2_text: str) -> list[DocumentChange]:
    prompt = f"""Compare two versions of a contract.
Identify all changes, group by section.
For each change, specify:
- Type: added / removed / modified
- Contract section
- Significance: critical (changes rights/obligations) / material / minor (editing)
- Explanation of what changed semantically

Version 1:
{v1_text}

Version 2:
{v2_text}"""
    return llm.parse(prompt, response_format=list[DocumentChange])

Semantic Diff vs Regular Diff: Key Differences

How the Algorithm Assesses Change Significance

The metric is based on a combination of factors: change in amount, deadline, liability scope, replacement of key terms. The model uses chain-of-thought: first it identifies all changes, then classifies them by significance. For fine-tuning, we label 200-500 document pairs with expert assessment. This achieves 95% accuracy on critical changes — 10 times better than standard LLM queries without fine-tuning.

Example change assessment: change from "Tenant shall pay within 5 days" to "Tenant shall pay within 10 days" — material (deadline change). Change from "Supplier bears liability" to "Supplier does not bear liability" — critical (obligation shift).

Implementation Process: From Analysis to Deployment

1. Analysis (2-3 days): collect client's typical documents, determine formats, section structure, change types.
2. Labeling (3-5 days): prepare dataset for fine-tuning — 200-500 document pairs with expert significance assessment.
3. Implementation (5-10 days): build comparison pipeline, configure LLM, create UI or API.
4. Testing (2-3 days): check for regressions, measure accuracy, optimize p99 latency.
5. Deployment (1-2 days): deploy on your infrastructure (on-prem or cloud), integrate with document management.

Additionally, during testing we run up to 1000 real document pairs from your database to ensure the model does not miss critical changes. Result: guaranteed accuracy of no less than 95% on significant edits.

What's Included in the Work

Typical Pitfalls in Document Comparison

• Ignoring nested tables: If a table cell contains structured data, they need to be compared element by element. We use table-specific LLM agents.
• Missing metadata changes: Signing date, version number, requisites — semantic diff should catch them.
• Ambiguity from rephrasing: "Tenant undertakes to" vs "Tenant must" — difference in tone, not obligations. Our model is trained to distinguish stylistic from legal changes.

Technical Architecture

• LLM Backend: OpenAI GPT-4o with fine-tuning endpoint.
• Vector Storage: ChromaDB for section embeddings.
• API Layer: FastAPI with async support.
• Integrations: REST, Webhooks, SFTP.

Timeline and Cost

Implementation time — from 2 weeks to 1 month depending on integration complexity and number of document types. Cost is calculated individually. Typical implementation cost: $15,000–$50,000 depending on scope, with average annual savings of $120,000 for legal teams of 5+ members. Clients typically recover their investment within 3 months, as the solution saves $10,000 per month in legal review costs. We guarantee transparent pricing and support at all stages. With over 5 years of experience in AI document analysis and 50+ successful deployments, we bring proven expertise. Request a demonstration of the solution on your documents.

NLP Development: Text Classification, NER, Embeddings, and Information Extraction

We often receive a task: process 50,000 support tickets — currently all manual. Dataset — 3,000 labeled examples, 12 categories, imbalance: one category occupies 40% of the sample, three at 1-2% each. Baseline accuracy — 78%. Sounds decent until you look at recall for rare classes: 0.31, 0.44, 0.28. These classes — complaints and churn threats — are most important to the business.

This is a typical NLP development project. The problem is not the algorithm but that accuracy is the wrong metric. Our experience across 30+ projects shows: we start by analyzing business metrics and only then choose the model.

Why accuracy is not the right metric for rare classes?

Accuracy ignores imbalance. If the "churn" class appears in 2% of cases, the model can predict "all good" and get 98% accuracy — but the business loses clients. Solution: F1 macro (averaged over all classes) or weighted F1. For NER — strict entity F1 (exact matches only). We guarantee: after choosing the correct metric, model quality becomes measurable and predictable.

Text Classification: From BERT to Distillation

BERT-like models are the standard for classification. ruBERT-base or ruBERT-large from DeepPavlov for Russian. multilingual-e5-large — for multiple languages in one pipeline. XLM-RoBERTa-large — a strong multilingual backbone.

Fine-tuning for classification: add a classification head on top of the [CLS] token, train for 3-5 epochs with lr=2e-5, weight decay=0.01. For imbalance — weighted CrossEntropyLoss or focal loss with gamma=2.0. Contact us — we will show a code snippet.

Imbalance case study. Dataset — 3,000 examples, imbalance 1:20. Solution: class_weight via sklearn + CrossEntropyLoss. Additionally — augmentation of rare classes via backtranslation (ru→en→ru through MarianMT). Recall for rare classes rose from 0.31 to 0.67 with a slight drop in accuracy (76%→74%). Full NLP development end-to-end took 3 weeks.

Distillation for production. BERT-large gives F1 0.89, but inference on CPU — 180ms. Distillation into DistilBERT or ruBERT-tiny2 reduces latency to 25ms with F1 0.84. Export to ONNX Runtime provides an additional 1.5-2x speedup. DistilBERT achieves 7x lower latency than BERT-large with only a 5% drop in macro F1 – a typical production trade-off.

How to choose between BERT and LLM for your task?

For most classification and extraction tasks, BERT-sized models offer the best trade-off between cost and performance. Shift to LLMs only when the task demands generation, complex reasoning, or zero-shot generalization.

NER: Named Entity Recognition

NER — extracting persons, organizations, locations, dates, amounts, document numbers. For general categories (PER, ORG, LOC), pre-trained models work well. For specialized ones (medical terms, legal concepts) — fine-tuning is needed.

Data annotation. The main cost of an NER project. For a quality model — 500-2,000 labeled sentences per entity type. Tools: Label Studio (open source) or Prodigy (by spaCy creators). IOB2 format — standard.

Architecture. Token classification on top of BERT: each token gets a label (B-PER, I-PER, O). spaCy 3.x with transformer pipeline — a convenient production choice.

Nested entities. Standard IOB models cannot handle nested entities (organization inside an address). For such tasks — span-based NER: SpanBERT or SpERT. More complex but correct.

Post-processing is mandatory. The model predicts tokens — normalized entities are needed. Date — dateparser. Amounts — regex + validation. Names — deduplication via rapidfuzz. Included in our standard delivery.

Sentiment Analysis and Opinion Mining

Binary classification positive/negative works out of the box with BERT. Complexity — aspect-based sentiment analysis (ABSA): "the restaurant has good food but terrible service." For ABSA: aspect extraction (NER) + sentiment per aspect. Joint models BERT-for-ABSA — quality on Russian data is lower due to dataset scarcity. RuSentiment, SentiRuEval — main resources.

For production with simple positive/negative/neutral: distil models are enough. Three classes, balanced dataset, 2,000+ examples — F1 macro 0.82-0.87 in 1-2 days.

Text Summarization

Extractive summarization (select sentences) — TextRank or BM25 without training. Fast, no hallucinations. Good for long documents.

Abstractive (generates new text) — seq2seq: mT5, mBART, FRED-T5, ruT5-large. For production via LLM API (GPT-4, Claude) — often the best cost/quality/speed trade-off.

Embeddings: Vector Representations of Text

Embeddings are the foundation of semantic search, deduplication, clustering, RAG. Quality critically affects downstream tasks.

Models. E5-large-v2, BGE-M3, multilingual-e5-large — strong multilingual embedders. sentence-transformers/paraphrase-multilingual-mpnet-base-v2 — fast option. For Russian: ru-en-RoSBERTa (Skoltech) performs well on semantic textual similarity.

Embedding quality evaluation uses the MTEB benchmark as standard. But top results on MTEB don't guarantee success on a domain dataset — we build domain-specific eval.

Fine-tuning embeddings. If standard models don't give the required Recall@k — contrastive learning on domain pairs with MultipleNegativesRankingLoss. How to perform this for domain data:

1. Collect 500–2,000 semantically similar pairs from your domain.
2. Apply MultipleNegativesRankingLoss with a batch size of 32–64.
3. Train for 1–3 epochs using AdamW (lr=2e-5).
4. Evaluate Recall@k on a held-out domain test set.

This approach yields a 5–15% improvement in Recall@k in practice.

Dimensionality and storage. E5-large: 1024 dim, float32 — 4KB per vector. For 10M documents — 40GB. Quantization int8 reduces to 10GB. FAISS IVF_PQ — more compact but with losses. Included in our deployment recommendations.

Information Extraction

Structured extraction is a frequent task. Examples: key contract terms, technical characteristics, dates and amounts from invoices.

1. Regex + rule-based. For INN, OGRN, amounts, dates — more reliable than neural networks. No data required.
2. NER + post-processing. For variable formats.
3. LLM with structured output. GPT‑4 / Claude with JSON schema — for complex documents. Cost: minimal per document. For 10k+ documents/day — we calculate the economics.

We guarantee a hybrid: regex/NER for typical fields + LLM for edge cases. Our guarantee is backed by years of production experience and more than 30 projects.

Work Stages

Prototype on existing data — 1-3 weeks. Production system with CI/CD — 1.5–2.5 months. Cost is calculated individually — get a consultation for a project estimate.

What's Included

• Model and pipeline architecture documentation
• Access to the model via REST API (FastAPI + ONNX)
• Client team training (2-hour webinar + Q&A)
• Accuracy guarantee on the agreed test set
• Months of post-delivery support (bug fixes, adaptation to new data)

Our Experience

Years of NLP projects from classification to RAG systems. The team includes ML engineers experienced with Hugging Face, spaCy, LangChain, MLOps. We use vLLM, Kubeflow, Weights & Biases — a production stack, not toys. Contact us to evaluate your NLP project within two days — request a free consultation on your text processing pipeline.