Legal departments spend up to 70% of their time on routine contract analysis and searching for relevant legal norms. We develop AI Legal Assistant — a digital worker that automates these tasks and integrates into your document workflow. The system reduces lawyer workload by 60% and operates 24/7. Average budget savings for a legal department amount to up to 60%: with an annual budget starting at 2 million rubles, this is over 1 million rubles per year. Below are the technical implementation details.
How AI Legal Assistant Analyzes Contracts
The core of the system is RAG on top of a regulatory database (Civil Code, Labor Code, Tax Code, industry-specific laws). Documents are split into paragraphs with 20% overlap to preserve context. We use text-embedding-3-large or multilingual-e5-large embeddings for Russian texts. The vector store is pgvector (PostgreSQL) or Weaviate for production loads. Hybrid search BM25 + dense retrieval with RRF ranking is twice as accurate as pure semantic search.
The document analysis module processes contracts, lawsuits, and corporate documents: structural extraction of parties, subject, terms, liability; identification of risky clauses; comparison with reference templates; generation of legal opinions.
Why Deploy a Digital Lawyer Now?
Typical risks the system identifies in contracts: unlimited liability without a cap, unilateral change of terms, absence of force majeure clauses, violation of antitrust laws. For each risk, it specifies the contract clause, references the legal norm, and suggests a revised version. Our clients report reducing contract review time from 4 hours to 20 minutes.
According to Gartner, a large share of enterprises will use AI assistants for legal work in the near future.
Stack and Architecture
| Layer |
Tools |
| LLM (primary) |
GPT-4o, Claude 3.5 Sonnet, or fine-tuned LLaMA for on-premise |
| Orchestration |
LangChain / LlamaIndex |
| Vector DB |
pgvector, Weaviate, Qdrant |
| Document processing |
Apache Tika, unstructured.io, pdfminer |
| OCR (scans) |
Tesseract 5, Azure Document Intelligence |
| Backend |
FastAPI + Celery |
| Frontend |
React + Lexical editor |
Fine-tuned LLaMA on on-premise shows p99 latency 2x lower than GPT-4o with comparable quality.
Learn more about RAG technology: Retrieval-Augmented Generation
Contract analysis pipeline:
Example pipeline
[Document upload]
→ [Text extraction: pdfminer / unstructured]
→ [Structural parsing: sections, articles, clauses]
→ [LLM extraction: parties, subject, key terms]
→ [Search in legal act base: relevant norms]
→ [Risk scoring: clause analysis against checklist]
→ [Opinion generation: Markdown / DOCX]
→ [Storage in vector DB for future search]
Key Modules
The legal opinion system is implemented via a chain of prompts: extraction chain → analysis chain → risk chain → recommendation chain. Each chain uses few-shot examples from real anonymized opinions to maintain a professional tone.
Risk identification: the model is trained on a checklist of typical risks and compares against best practices. For example, it recommends capping liability (limit of X annual salaries) rather than unlimited joint liability.
Jurisdiction handling: prompts explicitly specify the jurisdiction, and the RAG base is segmented geographically. Russian, Ukrainian, Belarusian law — different codes and case law. For international contracts, a comparative law module is added.
Integrations and Security
- 1C:Enterprise — bidirectional synchronization via REST API
- Diadoc / SBIS — receiving EDI documents for analysis
- Microsoft 365 — plugin for Word
- Telegram / Slack — notifications about legislative changes
Security: on-premise LLM deployment (LLaMA, Mistral) to prevent data exposure; encryption at rest (AES-256) and in transit (TLS 1.3); role-based access control; full audit log; automatic depersonalization for test environments.
Accuracy and Guarantees
| Metric |
Target |
| Extraction F1 |
>95% |
| Risk detection recall |
>90% |
| Hallucination rate |
<2% |
| User acceptance rate |
>80% |
Each citation of a legal act is verified by searching the database: if the norm is not found, the system marks the statement as unverified. We guarantee these metrics based on experience from 25+ implemented projects. Average budget savings for the legal department can reach up to 60%.
Implementation Process and Timelines
From 6 to 10 months depending on complexity. Stages:
Month 1–2: Build regulatory base, configure RAG, basic Q&A on legislation.
Month 3–4: Contract analysis module, integration with document workflow.
Month 5–6: Opinion generation, risk scoring, legislative monitoring.
Month 7–8: Integrations (1C, EDI), lawyer interface, load testing.
Month 9–10: Pilot with real users, quality iterations, production launch.
What Is Included
- Architectural documentation and stack description.
- Access to web interface and REST API.
- 2-day workshop for lawyers.
- 3 months of technical support after launch.
- Knowledge base updates when legislation changes.
Contact us for a project assessment. Get a consultation on AI Legal Assistant implementation. The investment pays off through reduced lawyer time and lower risks.
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.
| Model |
F1 macro |
Latency (CPU) |
Size |
| BERT-large |
0.89 |
180 ms |
1.3 GB |
| DistilBERT |
0.84 |
25 ms |
250 MB |
| ruBERT-tiny2 |
0.81 |
12 ms |
120 MB |
| DistilBERT + ONNX |
0.84 |
14 ms |
150 MB |
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:
- Collect 500–2,000 semantically similar pairs from your domain.
- Apply MultipleNegativesRankingLoss with a batch size of 32–64.
- Train for 1–3 epochs using AdamW (lr=2e-5).
- 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.
- Regex + rule-based. For INN, OGRN, amounts, dates — more reliable than neural networks. No data required.
- NER + post-processing. For variable formats.
- 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
| Stage |
Duration |
What's included |
| Data and metric analysis |
3-5 days |
Class distribution, text lengths, baseline |
| Baseline (TF‑IDF + LogReg) |
1 day |
Quick estimate of gap with deep models |
| Training and validation |
1-2 weeks |
k‑fold, early stopping, error analysis |
| Deployment (ONNX + FastAPI) |
1-2 weeks |
REST API, batching, monitoring |
| Documentation and training |
2-3 days |
Model card, API docs, team training |
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.
