AI Document Agent Integration with EDO (Diadoc, SBIS)

Electronic document management systems (EDO) store terabytes of documents but don't understand their content. They only transfer and sign. We change that: AI Workforce connects to your Diadoc or SBIS and processes documents like an experienced employee — read, analyze, fill in, make decisions. Accounting receives hundreds of invoices, acts, UPD daily. Each document must be manually opened, checked against the order, posted in 1С, and signed. Errors are inevitable: missed requisites, discrepancies in amounts, missed deadlines. AI Workforce solves these problems by automatically processing documents from receipt to posting. Reducing FTE costs by 70% through automation is a typical result after implementation. The project pays for itself in 4–6 months. Typical project cost starts from $50,000 with ROI in 4–6 months, saving 70% on document processing costs. Order a pilot in 2 weeks and see the effect.

We are ISO 27001 certified and have 5+ years of experience in EDI automation. With 20+ successful implementations, we guarantee results.

We build an agent layer on top of your infrastructure. AI agents become a participant in the document flow, but fully automated. The agent layer includes five types of agents: document reception, classification, data extraction, validation, and actions (signing, routing). Each agent solves its task using fine-tuned models and RAG pipelines for document workflow automation. This modular architecture allows flexible configuration for any business processes.

After connecting to the EDO API, the system starts monitoring incoming documents. A new document enters the receiver agent, which downloads its content. Then the classifier determines the document type and routes it to the appropriate extractor. The extracted data undergoes validation, and the action agent decides: sign, reject, or send for manual review. The entire cycle takes seconds. Our AI workforce integrates with Diadoc, SBIS, and 1С for seamless document workflow automation. Using RAG and fine-tuned BERT, we perform 3-way matching and data extraction.

How AI Workforce connects to EDO?

[EDO system (Diadoc / SBIS / 1С-EDO)]
    ↕ REST API / SOAP / Webhook
[AI Integration Layer]
    ├── Document Receiver Agent
    ├── Classification Agent
    ├── Extraction Agent
    ├── Validation Agent
    └── Action Agent (signing, rejection, routing)
    ↕
[Internal systems: 1С, ERP, CRM, DB]

Diadoc: API Integration

Diadoc provides REST API with OAuth 2.0. Key endpoints:

• GET /v1/GetNewEvents — get new documents (polling or webhook)
• GET /v1/GetDocument — download document body (XML for formalized, PDF/DOCX for non-formalized)
• POST /v1/PostMessage — send a signed document
• POST /v1/Delete — reject with comment

For signing, we use CryptoPro DSP API or a local crypto provider. The agent calls signing through a separate secure service, does not store keys.

SBIS: Integration via WebAPI

SBIS uses JSONRPC API (SBIS WebAPI). Authentication via SID session. Main methods:

• СБИС.ЗаписатьДокумент — create and send
• СБИС.СписокДокументов — get list with filtering
• СБИС.ПрочитатьДокумент — get content

Specific to SBIS: documents often come in SBIS-XML format, requiring a custom parser. We place an intermediate converter to unified JSON. Our custom parser is documented in our integration guide.

Why AI is faster than manual processing?

Compare with a full-time document clerk:

Which documents does AI process automatically?

Formalized invoices, UPD, certificates — accuracy 97–99%. Non-formalized PDF and Word — accuracy 90–95% after custom fine-tuning. All exceptions are routed to humans. For non-formalized documents, we use fine-tuned BERT or prompt-based GPT-4o, achieving quality sufficient for industrial operation. Our extraction pipeline achieves 95% accuracy after 2 weeks of training on your data.

How do agents process documents?

Classification and Routing

The first agent is the classifier. It determines the document type and route:

The classifier is trained on your document corpus (fine-tuned BERT or prompt-based GPT-4o). Accuracy 97–99% for formalized, 90–95% for non-formalized.

Three-Way Matching

Key task of AI Workforce in finance is automatic reconciliation of order, delivery note, and invoice. The agent checks items, quantity, price, totals. When discrepancy exceeds threshold → flag for manual review; when aligned → automatic signing and payment initiation. Uses 3-way matching with thresholds:

• Amount: ±0.5% (rounding tolerance)
• Quantity: 0% (exact match)
• Items: fuzzy matching with 85% threshold

Integration with 1С

Two-way synchronization via 1С REST API (oData) or COM objects:

• From 1С to EDO: automatic creation and sending of outgoing documents
• From EDO to 1С: creation of documents based on received ones (UPD → Goods receipt, Certificate → Service receipt)

For 1С:ERP, we use the "Electronic Document Exchange" subsystem with an extension for AI validation before posting.

Exception Handling and Human Control

Not everything goes through automatically. The system routes exceptions:

• New counterparty → verification via Federal Tax Service/Unified State Register of Legal Entities → manual approval
• Amount above threshold → mandatory manual authorization
• Data discrepancy → notification and blocking
• Expiration → automatic renewal or escalation

Monitoring and SLA

Key metrics in production:

• Straight-through processing rate — share of documents without manual intervention: target 70–85%
• Processing latency — up to 5 minutes for standard documents
• Extraction accuracy — >98% for formalized
• False positive rate — <5%

Security and Compliance

• Digital signature is stored in HSM or secure crypto service
• AI agent has no direct access to keys
• Full audit trail for tax authorities
• Compliance with Federal Law No. 63-FZ on Electronic Signatures

Implementation Timeline

Brief Implementation Scheme

1. Audit of current document flows and preparation of API access to EDO.
2. Deployment of agent layer: receiver, classifier, extractor.
3. Setup of 3-way matching and 1С integration.
4. Pilot launch for 2–3 weeks on real data.
5. Optimization of thresholds and exception routes.
6. Production launch with full monitoring.

Common integration issues: Diadoc API limits (no more than 100 requests per minute per session), SBIS-XML and XML UPD format differences (intermediate converter required), incorrect item settings in 1С (leading to false discrepancies), lack of webhook notifications from some operators (polling must be configured). We address all these nuances during the audit phase.

What's included

• Integration documentation (flow diagrams, agent configurations, API call descriptions)
• Configured AI agents: receiver, classifier, extractor, validator, action agent
• UI for exception handling and manual control
• Operator training (2–3 days)
• Support for 2 weeks after production launch
• Source code transfer (if required) for in-house maintenance

After the pilot, you get a working prototype ready for industrial scaling. Get a consultation for more details.

We have implemented AI Workforce at 20+ large enterprises. Our accumulated experience allows us to predict bottlenecks and configure the system in 12 weeks turnkey. Contact us to evaluate your document pipeline.

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.