AI Document Search Implementation for Archives (Document Search)

AI Document Search Implementation for Archives (Document Search)

File systems are powerless against semantic search: a contract with auto-renewal cannot be found if the query does not contain the exact phrase. Managers spend hours browsing folders; lawyers miss deadlines because of lost contracts. We implement AI systems that understand the query semantically, not just search for a substring. Result: seconds instead of hours. For a typical archive of 50,000 documents, search time drops from 15 minutes to 30 seconds.

How Hybrid Search Works

Hybrid search combines two approaches: semantic (embeddings) and lexical (BM25). Semantic captures synonyms and context: a query "extend contract with Gazprom" will find the phrase "prolongation of agreement with PJSC Gazprom". Lexical ensures exact match search for numbers, dates, amounts. We mix results via RRF (Reciprocal Rank Fusion) and rerank with a cross-encoder. Final accuracy (NDCG@5) on a test collection of 10,000 documents is 0.89. The hybrid approach boosts recall by 20% compared to pure embeddings and yields 1.5x higher nDCG@5 than BM25.

Indexing the Archive

Each document entering the archive undergoes processing:

1. Text extraction: pdfminer (PDF), python-docx (DOCX), unstructured.io (supports 30+ formats).
2. Structuring: split into chunks of 512 tokens with 128 overlap + preserve metadata (section, page, creation date).
3. Embeddings: text-embedding-3-small (OpenAI, 1536-dimensional) or cointegrated/rubert-tiny2 (384-dimensional, on-premise). Model choice affects latency: GPU inference takes ~20 ms per chunk.
4. Indexing in Qdrant or pgvector with HNSW index for fast search over 1M+ vectors (latency p99 < 300 ms).
5. Extracting structured metadata: document type, parties, dates, amounts — using an NER model (spaCy + fine-tuned on your data) and writing to a relational DB.

Why Cross-Encoder Reranking?

After obtaining the top-K from hybrid search, we apply a cross-encoder model (e.g., cross-encoder/ms-marco-MiniLM-L-6-v2) that pairwise assesses the relevance of each document to the query. This adds 50–100 ms to latency but improves first-page accuracy by 15–20%. In practice, users rarely scroll to the second page.

Faceted Search

Additional filters for precise search:

Facets combine with the semantic query: search "lease contracts over 1 million" and immediately see only active ones.

What Is Conversational Search?

We implemented a conversational mode: the system progressively refines search parameters — counterparty, period, document type — and converts the dialogue history into a structured query to the storage. An LLM (GPT-4o or LLaMA 3 70B) converts the conversation into filter parameters. No more remembering column names in Excel or clicks in CRM. We implemented this scenario for five legal entities with archives of 50,000+ documents — search time dropped from 15 minutes to 30 seconds.

Comparison of Search Approaches

Hybrid approach gives the best balance: 20% more recall than pure embeddings and 35% higher nDCG@5 than BM25.

What Is Included

Our implementation includes:

• Indexing pipeline (Python + Apache Airflow) for your storage.
• Vector DB (Qdrant) with tuned index parameters.
• API endpoints for search (REST/gRPC) with facet support.
• Web interface with search bar and card results.
• Operations documentation and training for your engineers (2-day workshop).
• One month of technical support after launch.

Our Experience and Guarantees

We have been doing AI search for over 5 years and have completed 50+ projects for banks, logistics companies, and government. Our team includes certified machine learning specialists (MLflow, Kubeflow). We guarantee: the system will find what you are looking for with at least 90% accuracy on a test sample.

Timeline and Cost

Implementation timeline for a pilot project: 3 to 6 weeks depending on archive size and customization required. Pricing is calculated individually based on document count, number of metadata fields, and required SLA. Contact us to receive a commercial proposal. Order a pilot project on a test sample — verify effectiveness before full deployment. Get a consultation on implementation right now.

Reference: BM25 — a classic ranking function for text relevance evaluation.

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.