Dialog Management for AI Bots: Design and Implementation

Most AI bots lose context after two or three messages. Clients repeat themselves, the bot responds nonsensically, and the dialogue hits a dead end. The root cause is the lack of a dialog management system. As AI/ML engineers, we solve this by designing a component that remembers history, manages state, and selects appropriate actions. End-to-end, with documentation and support. This is not just an abstract module—it's the core that determines user retention and CX effectiveness.

How Does Dialog Management Solve Context Loss?

Dialog management stores the conversation state: current intent, filled slots, message history, and based on that decides which response to give. Without it, a bot cannot carry a coherent conversation longer than a few exchanges. In production systems, dialog management is a critical component that determines user experience quality and support load. For example, if a user writes "I want to order pizza," the manager must remember that the intent is ordering and sequentially collect slots: size, toppings, address. It should not re-ask for already filled data.

How to Choose a Dialog Management Model?

• Finite State Machine (FSM) — explicit states and transitions. Reliable, predictable, but for complex scenarios the number of states grows exponentially.
• Frame-based — sets of forms with slots (booking, ordering).
• LLM-based — the model decides based on history, maximally flexible but less predictable.
• Hybrid (production-best) — FSM for critical paths (payment, authorization), LLM for free-form dialogue.

The key criterion is predictability for critical paths. For financial transactions or medical data, FSM is needed; for general conversation, LLM. We always recommend hybrid.

Why Is Hybrid Architecture the Production Standard?

Hybrid architecture reduces wrong answers by 30–40% compared to a pure LLM solution—that is 1.5 times more effective and 3x more reliable for critical paths. In one fintech project, we implemented such a scheme: FSM handled 80% of traffic (short balance inquiries, transfers), LLM handled 20% (complex questions, complaints). This reduced p99 latency from 1500 ms to 120 ms and cut operator escalations by 35%, resulting in annual savings of $50,000. Contact our engineers to assess your project.

How to Store Dialog State?

Dialog state must be persisted between sessions and after bot restarts. For active sessions we use Redis with TTL (30 minutes). State is serialized to JSON and stored by conversation_id. For long-term history and analytics we use PostgreSQL. This approach ensures dialog recovery after any failure and low latency. A sample state structure is shown below.

@dataclass
class DialogState:
    conversation_id: str
    user_id: str
    current_intent: str | None
    filled_slots: dict
    dialog_history: list[DialogTurn]
    context: dict          # business context (user profile, session)
    flow: str              # "main_menu" | "booking" | "support" | "handoff"
    pending_action: str | None  # expected confirmation/input

What Is Policy and How to Implement It?

Policy is the algorithm that selects the bot's next action. Rule-based: if-else tree—transparent but limited. Learned policy (Rasa Core) uses a neural network on dialogue stories—flexible but requires data. LLM policy: language model chooses an action from a tool set. In production we use a hybrid: rule-based for critical paths, LLM for resolving ambiguities.

class DialogManager:
    def process_turn(self, state: DialogState, user_input: str) -> BotAction:
        # Update history
        state.dialog_history.append(DialogTurn(role="user", text=user_input))

        # Detect intent
        intent = self.intent_detector.detect(user_input)

        # Decide next action
        if self.should_escalate(state, intent):
            return HandoffAction(reason=EscalationReason.USER_REQUEST)

        if state.flow == "booking" and state.pending_action == "confirm":
            return self.handle_booking_confirmation(state, user_input)

        # Update slots
        state.filled_slots = self.slot_filler.update(user_input, state.filled_slots)

        # Select next action
        return self.policy.select_action(state, intent)

What Performance Metrics Will You Achieve?

Development Process and Timeline

1. Analysis: collect dialogues, identify scenarios.
2. Design: draw FSM diagram, define slots, escalation rules.
3. Implementation: write DialogManager, integrate with NLP pipeline (Rasa, LLM API).
4. Testing: dialogue simulation, A/B tests.
5. Deployment: containerization, monitoring (latency, accuracy).

Timeline: 3 to 6 weeks depending on scenario complexity.

What's Included in the Deliverable

• documentation of states and transitions,
• DialogManager module code with tests,
• integration with Redis/PostgreSQL,
• training for the client's team,
• 3-month warranty support.

Our team's experience—50+ AI bots in production, 10+ years in NLP and MLOps. Contact us for a project assessment. We guarantee that the dialog management will be robust to edge cases and handle loads up to 10,000 requests per minute.

Finite-state machine (FSM) — a mathematical model for describing system behavior through a finite number of states.

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.