AI classification of incoming documents: implementation and tuning

Implementing AI classification of incoming documents by type

The incoming correspondence department processes 500+ documents per day. Employees manually determine the type — whether it's an invoice or a contract — and then enter it into the system. Errors occur in 15% of cases, each document takes 3–5 minutes. We know this pain, so we build automatic classifiers that work with 97–99% accuracy and save up to 80% of time on routine routing. For example, for a large logistics operator, we implemented a classifier that processes 2000+ documents daily with 98.5% accuracy, completely replacing manual sorting.

Key problems solved by AI classification

Low accuracy for similar documents

An invoice and a waybill often have identical fields and structure. A text classifier will err in 5–10% of cases. A multimodal approach — text + tables + file metadata — reduces error to 1–3%.

Unknown types

A system without the UNKNOWN class sends an unfamiliar document to the nearest category with low confidence. This disrupts business processes. We allocate a separate class for manual processing and log all cases for retraining.

Integration with existing ECM/ERP

API layer based on FastAPI or GraphQL, support for REST, SOAP, gRPC. The classifier can be easily embedded into the processing pipeline — from scanning to loading into 1C, SAP, or DocuWare.

How the multimodal classifier works

Our stack: PyTorch, HuggingFace Transformers, LangChain for chains, ChromaDB or Qdrant for storing embeddings. Models — rubert-tiny2 for Russian or multilingual-e5-large for multilingual document flow. Deployment via Triton Inference Server with INT8 quantization support — p99 latency < 100 ms.

def classify_document(file_path: str) -> DocumentClass:
    features = {}
    # Text features
    text = extract_text(file_path)
    features["text_class"] = text_classifier.predict(text[:2000])
    # Structural features
    features["has_tables"] = detect_tables(file_path)
    features["page_count"] = get_page_count(file_path)
    features["filename_hint"] = extract_filename_hint(file_path)
    # Document metadata
    features["creation_date"] = get_document_metadata(file_path).get("created")
    # Ensemble decision
    return ensemble_classifier.predict(features)

Case study: classifying 500K documents per month

A large retailer (our client) implemented a processing system for waybills and acts. Before implementation — 4 employees on manual sorting, 85% accuracy. After — an AI classifier based on RuBERT fine-tuned on 20K labeled documents. Accuracy: 98.5% for waybills, 97.2% for acts. Processing time per document — 0.7 s. Result: 3 out of 4 employees were reassigned to quality control, the cost of processing one document decreased by tens of times.

Why multimodal approach is better

The multimodal approach is 3 times more accurate than pure-text on documents with similar structure.

Why invest in model fine-tuning?

Fine-tuning on your data boosts accuracy by 5–10% compared to an out-of-the-box model. For a client handling 1000 documents per day, the savings from reduced rework cover the fine-tuning cost within the first month.

Implementation process

What's included in the result

• Ready classifier model (fine-tuned for your taxonomy)
• API for integration (OpenAPI specification)
• Docker images for deployment (CPU/GPU)
• Operation manual and operator training
• Guaranteed accuracy of at least 95% on the test set
• 3 months of post-launch support

Typical mistakes in self-implementation

• Ignoring layout features. A simple BERT classifier confuses an invoice and a payment order if the texts are similar. Add table embeddings and page counts — accuracy will increase by 10–15%.
• Missing the UNKNOWN class. Always provide a fallback for unknown types. Without it, a classification error breaks the processing chain.
• Insufficient labeled data. Fine-tuning requires at least 200–500 examples per class. Less leads to high variance, more is better.

Timeline and budget

Implementation timeline — from 4 to 12 weeks depending on taxonomy complexity and integration. The budget is calculated individually after analyzing the document flow. Contact us — we will evaluate your project and propose a scenario with a guaranteed result. Request a consultation to get savings of up to 80% of time on manual sorting.

Experience: 10+ years in AI/ML, 80+ document classification projects, certified specialists in PyTorch and HuggingFace.

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.