Modern AI Agent Assist for Accelerating Knowledge Base Search in Support

Improve Support Efficiency with AI Agent Assist for Knowledge Base Search

Support agents spend up to 30% of their time searching for the right answer in the knowledge base. Every extra second increases Average Handle Time (AHT) and lowers Customer Satisfaction (CSAT). Traditional keyword search delivers irrelevant results: synonyms and context are ignored, forcing agents to scroll through dozens of articles. Studies indicate that deploying an AI assistant in support reduces AHT by 25–40%.

Our AI-powered support system analyzes conversations in real time and retrieves relevant articles, instructions, and response templates. We have deployed such solutions for a dozen companies handling 500 to 10,000 daily tickets, achieving an average AHT reduction of 30–40%. A critical requirement is p99 latency below 1 second, achieved by caching embeddings, precomputing indices, and using GPU inference (T4 or A10G). For a team of 50 agents handling 200 tickets daily, this translates to annual labor savings of over $150,000.

How Answer Suggestion Works

Every new customer message triggers a chain of actions:

1. Intent detection — identify the topic (e.g., "return product" or "password reset").
2. Parallel search across three channels:
   • Vector search on knowledge base embeddings (using text-embedding-3-small with dimension 1536 and HNSW indexing for fast approximate nearest neighbor);
   • BM25 search on FAQs and templates (sparse retrieval);
   • Search through resolved tickets with the same intent.
3. Re-ranking — an ensemble of a lightweight CatBoost model and a document-freshness rule, plus a cross-encoder for precision.
4. Brief summary generation (optional, via LLM for complex cases).

Why Vector Search Outperforms BM25 Alone?

BM25 works well with exact keywords but fails on synonyms and context. Vector search (cosine similarity on embeddings, a form of dense retrieval) finds semantically close documents even if wording differs. For example, a query "how to change my password" retrieves the article "account reset".

In practice, the best results come from a hybrid approach: combining BM25 and vector search with weights that we tune via cross-validation on your historical data. Hybrid search provides 85% recall@5, which is 25% better than BM25 alone.

Data Sources Used

• Knowledge base: Confluence, Notion, internal Wiki — articles and instructions.
• Ticket archive: resolved tickets tagged "successful" — practical solutions.
• Response templates: ready-made phrases for typical cases (up to 80% of tickets).
• Product documentation: technical specs, API docs.

All sources are connected via REST APIs or direct integration using ETL pipelines (Apache Airflow). The overall pipeline is a Retrieval-Augmented Generation (RAG) architecture.

Search Method Comparison

Impact on Support Metrics

Operator Panel Interface

A side panel shows 3–5 most relevant articles with brief summaries. One click inserts the full content into the reply box, editable. For templates, an "Insert as-is" button is available.

Adoption metrics we track:

• Suggestion acceptance rate — target >55%.
• Modification rate — how often the agent edits the suggestion.
• AHT before and after — reduction of 30–40%.
• CSAT — answer quality does not suffer; often improves.

Learning from Accepted and Rejected Suggestions

Every acceptance or rejection is a signal to retrain the ranker. We use ranking fine-tuning to adapt to your data. We run A/B tests with different algorithms on a subset of agents. Over 3–6 months, the acceptance rate grows from 40% to 60–70% as the system learns from your data.

What's Included in the Work

1. Audit of your current knowledge base: evaluate quality and completeness, remove duplicates.
2. Build data pipeline: extraction, cleaning, vectorization (using sentence-transformers or OpenAI embeddings).
3. Select and tune the ranker: hybrid BM25 + embeddings, trained on your history.
4. Integration with CRM/Helpdesk: Zendesk, Freshdesk, Bitrix24, or your system via API.
5. UI component for the agent panel (web widget or extension).
6. A/B testing and calibration.
7. Documentation and training for the support team.

Based on 50+ AI solution deployments, the project from audit to first hypothesis takes 4–8 weeks. Full rollout with learning takes up to 3 months. With 5+ years of experience in AI support solutions, we have deployed over 50 projects. Typical investment ranges from $20,000 to $40,000 for a mid-size company, with ROI achieved within 3-6 months.

Contact us to discuss your knowledge base and get a preliminary project assessment. Request an audit of your current knowledge base — we guarantee an acceptance rate above 50% after the first two weeks of use. Try a demo version to see the effect on your own 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.