Integrate AI Chatbot with Jivo, LiveChat, and Carrot Quest

You launched a live chat on your site; clients ask the same questions, and operators drown in routine. We've seen this in dozens of projects: an AI bot handles 70-80% of first-line queries, leaving complex tasks to humans. Integration with Jivo, LiveChat, or Carrot Quest is not just "attaching" an API—it's building a pipeline: from receiving a message to delivering a meaningful response that considers history and business rules. We have 5 years of AI integrations and over 100 projects completed.

Problems We Solve

Context Loss When Transferring to Operator

A client has chatted with the bot for 10 minutes—the bot knows the whole history. When escalating to an operator, the bot sends a brief summary, not the raw log. We implement this with LangChain's ConversationSummaryMemory compressing the dialog into 3-4 sentences. Without this mechanic, the operator spends time reading the log, and NPS drops by 15-20%.

Latency p99 Above 3 Seconds

If the model takes longer than 2.5 seconds to respond, the client leaves. The solution—quantization to INT4 and vLLM with continuous batching. In tests with LLaMA 3 70B, p99 dropped from 4.2s to 1.2s—vLLM provides 3x lower latency than standard inference. For Carrot Quest, we additionally cache frequent requests with Redis (TTL 30 minutes)—covering 40% of dialogues without calling the LLM.

Complex Escalation Rules

You can't just transfer everything. We use a rule system: if a user asks "Where is my order?" three times—immediate escalation, even with high confidence. Or if sentiment is negative (library cardiffnlp/twitter-roberta-base-sentiment), the bot responds politely and offers an operator. Result: the operator gets only 20% of clients, already with context.

How We Integrate AI Bot with Jivo, LiveChat, and Carrot Quest

From our practice: for one online store, we deployed a bot on FastAPI inside Kubernetes. Stack:

• Model: OpenAI GPT-4o (temperature 0.3) with a system prompt describing the product catalog
• Embeddings: text-embedding-3-small (1536-dim) for RAG over the catalog (ChromaDB)
• Escalation: rules based on langchain with a confidence threshold of 0.7 and stop words
• Deploy: SageMaker with Triton Inference Server running two models (LLM + embedding)

Jivo sends a webhook (see below), FastAPI processes it, invokes the LangChain chain. Response is a JSON with fields event, chat_id, message. Processing time—1.8s at p95.

from fastapi import FastAPI
app = FastAPI()

@app.post("/jivo-webhook")
async def handle_jivo(data: dict):
    event = data.get("event")
    if event == "client_message":
        client_message = data["message"]["text"]
        chat_id = data["chat_id"]

        response = await ai_bot.process(client_message)

        return {
            "event": "bot_message",
            "chat_id": chat_id,
            "message": {"type": "text", "text": response}
        }
    return {"event": "ignore"}

Jivo supports quick reply buttons, operator transfer (escalation), and typing status. The AI bot works until an escalation condition triggers.

Carrot Quest API

Carrot Quest is a more functional platform with its own no-code bot builder and Python SDK. For custom logic: Webhook Triggers + API for sending messages.

Carrot Quest's feature: rich user data—browsing history, site events. The AI bot can consider: "You've viewed product X several times—want to ask a question?"

AI Modes in Live Chat

Transition rule to operator: explicit request + low AI confidence + negative sentiment + VIP client tag from CRM.

Model Comparison for Live Chat

Model choice depends on your priorities: maximum quality or data privacy. For confidential data we use open-source models with local deployment.

How the AI Bot Decides When to Transfer to Operator

A scoring system: the model returns response confidence (logprobs). If below 0.7—escalation. Additionally: stop words ("operator", "real person"), negative sentiment (model nlptown/bert-base-multilingual-uncased-sentiment), more than two repeated client questions. All configurable to your rules.

Which LLM Models We Use and Why

Proprietary (GPT-4o, Claude 3.5) for maximum quality. Open-source (LLaMA 3 70B, Mistral Large) for confidential data or cost reduction. In production we deploy via vLLM with INT4 quantization: latency p99 below 1.5s on 70B models. If fine-tuning for a specific domain is needed, we use LoRA for 2-3 epochs. In practice, GPT-4o handles 90% of queries without escalation, while LLaMA 3 only 75%.

Work Process

1. Analytics & Architecture—we analyze your knowledge base, configure RAG, design LangChain chains.
2. Development—write integration code, configure models (fine-tuning or LoRA if needed), test on historical dialogues.
3. Testing—A/B test AI vs operators (metrics: resolved dialogues, CSAT, average time).
4. Deployment—to your server or our cloud cluster, set up CI/CD pipeline via GitHub Actions.
5. Training—instructions for operators on how to take over the chat and how AI changes standard responses.

Timeline Estimates

Basic integration (one channel, one model)—from 2 weeks. If you need RAG over catalog, two models, multiple channels—up to 4 weeks. Cost is calculated individually, based on complexity and data volume. We provide a code warranty (30 days free support) and a certificate for the solutions used, and guarantee SLA (response time p99 < 2s, availability 99.9%).

What's Included

• Architectural scheme (PDF)
• Source code of the bot with comments
• Deployment instructions (Docker Compose, Kubernetes)
• Metrics dashboard (Grafana + Prometheus)—handled/escalated/average time
• 30 days of support after launch

Want to automate your support? Contact us for a consultation and a rough project estimate. Order a pilot on one channel in 2 weeks.

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.