Implementing Semantic Search for Text Documents

Standard full-text search (BM25) fails with synonyms, paraphrasing, and typos. A query like "how to increase team motivation" finds documents on "employee management methods" without a single word match. This is a fundamentally different architecture requiring vector representations and ANN indexes. We implement such systems turnkey, from data audit to production deployment. More about the concept can be read in the article about semantic search.

Semantic Search Architecture

Bi-encoder — the main working mode: separate models encode the query and documents into a common vector space. Search reduces to finding nearest vectors via ANN (Approximate Nearest Neighbor). Cross-encoder works at the reranking stage: it takes a query+document pair and outputs an accurate relevance score. It is slower (O(N) vs O(log N)) but gives maximum precision. The combination bi-encoder (retrieve) + cross-encoder (rerank) is the production standard. According to Reimers & Gurevych, this duo significantly outperforms each method individually.

Compare the main embedding approaches:

Which embedding model to choose?

For Russian language we use cointegrated/rubert-tiny2 as a baseline — fast, compact (312-dim vector). For maximum quality — intfloat/multilingual-e5-large or sbert-base-ru-mean-tokens (768-dim vector). Fine-tuning on your data gives a 5-10% NDCG boost. We select the model based on corpus size and latency requirements (p99 up to 100 ms).

from sentence_transformers import SentenceTransformer, CrossEncoder

# Bi-encoder
bi_encoder = SentenceTransformer("cointegrated/rubert-tiny2")
# For better quality: "intfloat/multilingual-e5-large"

# Cross-encoder
cross_encoder = CrossEncoder("cross-encoder/ms-marco-MiniLM-L-6-v2")
# For Russian: "DiTy/cross-encoder-russian-msmarco"

Qdrant vs FAISS: Which to choose for production?

Qdrant — production-grade, supports hybrid search, filters, replication. We recommend it for enterprise solutions. FAISS — in-memory index, requires no separate service. Ideal for prototypes and small corpora (<1M vectors).

Example indexing in Qdrant:

from qdrant_client import QdrantClient
from qdrant_client.models import Distance, VectorParams, PointStruct

client = QdrantClient("localhost", port=6333)
client.create_collection(
    collection_name="documents",
    vectors_config=VectorParams(size=312, distance=Distance.COSINE),
)

embeddings = bi_encoder.encode(documents, batch_size=64, show_progress_bar=True)
client.upload_points("documents", [
    PointStruct(id=i, vector=emb.tolist(), payload={"text": doc})
    for i, (emb, doc) in enumerate(zip(embeddings, documents))
])

What does hybrid search give?

Semantic search + BM25 outperform each method individually. BM25 catches exact matches (numbers, unique terms), while embeddings capture semantic proximity. Hybrid approach improves NDCG@10 by 2-3 times compared to pure BM25. We use RRF (Reciprocal Rank Fusion) to merge results.

from rank_bm25 import BM25Okapi
bm25 = BM25Okapi([doc.split() for doc in corpus])
semantic_scores = cosine_similarity([query_emb], doc_embeddings)[0]

def rrf(bm25_ranks, semantic_ranks, k=60):
    scores = {}
    for rank, idx in enumerate(bm25_ranks):
        scores[idx] = scores.get(idx, 0) + 1/(k + rank)
    for rank, idx in enumerate(semantic_ranks):
        scores[idx] = scores.get(idx, 0) + 1/(k + rank)
    return sorted(scores, key=scores.get, reverse=True)

Search Quality Evaluation

• NDCG@10 — normalized discounted cumulative gain. Takes order into account.
• MAP — mean average precision across all queries.
• MRR — reciprocal rank of the first relevant result.

Evaluation requires qrels (a set of queries with relevance annotations). We automate its creation: LLM generates questions for each document, the document itself is the "golden" answer. This yields a representative sample for metrics.

Implementation Process and Timeline

1. Data audit: volume, format, language, specific terms. Preprocessing includes cleaning, lemmatization, and chunking (chunk size ~512 tokens with overlap 128).
2. Architecture selection: bi-encoder + cross-encoder, hybrid, custom model. For large corpora (>10M documents) we use Qdrant clustering with sharding.
3. Pipeline development: chunking, embedding, indexing with latency p99 monitoring.
4. Tuning and deployment: Qdrant cluster (Helm charts), A/B testing, canary rollout.
5. Documentation handover, team training (2 sessions of 2 hours), 3-month warranty.

Timeline: from 2 weeks for a prototype, from 2 months for a production solution. Cost is calculated individually — contact us for a free assessment.

What is included in the result

• Detailed architectural documentation.
• Source code of the pipeline with comments.
• Integration with your infrastructure (Elasticsearch, databases, clouds).
• Deployment with Helm charts and CI/CD.
• Team training (2 sessions, 2 hours each).
• Support during industrial operation (1 month).

Why trust us

We are trusted due to 5+ years of experience and 20+ completed projects. All solutions are covered by unit tests and benchmarks. Our engineers are authors of open-source tools for embeddings and ANN. Get a consultation on your project — we will assess the task within 1 day. Order a pilot project to see results on your data.

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.