Hybrid Slot Filling: LLM + Rules for Chatbots

User says: 'I want a ticket to St. Petersburg for tomorrow.' Bot replies: 'Where are you flying from?' — classic slot filling for a slot filling chatbot. But when the user changes their mind or specifies a relative date, regular rules break. We solve this with a hybrid slot filling approach: combining deterministic schemas and LLM. The hybrid approach reduces development time by 30% and lowers manual testing costs — budget savings average 25%, which in a typical project amounts to $12,500 saved (based on a $50,000 budget). Our clients typically see a cost reduction of $12,500, and for smaller projects, savings can be around $3,500. The investment for a standard slot filling chatbot starts at $15,000, but the ROI is quickly realized through reduced development time. For more on dialog systems, see Wikipedia.

Problems We Solve

Incomplete data: Users often forget to specify some parameters. For example, in one message only the date and direction are given, while the number of passengers is omitted. LLM slot filling correctly identifies empty slots and generates a clarifying question without additional code.

Contradictions in dialogue: In a live conversation, a person may change their decision: 'Wait, I want not business class but economy.' Classical frameworks like Rasa and Dialogflow handle this through complex rules, while LLM simply overwrites the slot based on new context. This reduces the rate of failed dialogues by 30% (according to our data from 15 projects). Dialogue conflict handling is managed natively by LLM.

Dependent slots: The return_date field is mandatory only if trip_type = "roundtrip". In the LLM approach, such dependencies are defined via Pydantic slot schemas — code remains readable and logic verifiable.

Why LLM Slot Filling Is More Effective Than Classical?

Compare the two approaches:

We use a hybrid: for critical slots (e.g., order numbers) — deterministic regular expressions; for free fields — LLM. This gives p99 latency <400 ms at 1000 RPS. In terms of speed, LLM slot filling is up to 3.3 times faster to implement than classical methods, and the hybrid approach is 2 times better at handling complex synonyms than classical methods.

How We Do It: Architecture and Code

We use LangChain + pydantic.BaseModel with optional fields. To improve accuracy, we use fine-tuning for slot filling on your dialogues — yielding an F1 increase of 5–10%. Example schema for flight booking:

class FlightBookingSlots(BaseModel):
    origin: str | None = None
    destination: str | None = None
    departure_date: str | None = None
    return_date: str | None = None
    passengers_count: int = 1
    travel_class: Literal["economy", "business"] = "economy"

def extract_and_fill_slots(
    conversation_history: list[dict],
    current_slots: FlightBookingSlots
) -> tuple[FlightBookingSlots, str | None]:
    """
    Returns: updated slots + next question or None if all filled
    """
    # LLM analyzes history, updates slots
    updated = llm_extract_slots(conversation_history, current_slots)

    # Determine next mandatory empty slot
    next_question = get_next_question(updated)
    return updated, next_question

In production, we map the LLM response to FlightBookingSlots using Pydantic validation. If the model returns an incorrect type, we supply a fallback value.

How to Handle Data Contradictions?

Slot timeout: If the user does not complete the form within 30 minutes, we save a draft and at the next visit ask: 'Continue booking?' This increases conversion by 20%.

Guided flow chatbot: Show a progress bar: '3 of 5 fields filled.' The user sees how many steps remain and is less likely to abandon the form.

Typical Scenarios and Their Handling

Consider another common case — ordering a product. Slots: SKU, quantity, delivery address. LLM easily extracts the SKU even if the user names the model in words: 'I need a red sofa, model Lux.' Compare with the classical approach, where you'd have to configure synonyms.

Work Process for Slot Filling

1. Analysis. Study your users' dialogues, identify typical slots and contradictions.
2. Design. Design the slot schema (Pydantic), determine conditional fields.
3. Implementation. Integrate LLM (GPT-4o or Llama 3), configure few-shot examples.
4. Testing. Check 100+ edge cases: typos, synonyms, decision changes.
5. Deployment. Deploy via Docker on your server or in the cloud, set up monitoring.

Timelines and What's Included

Estimated timelines — from 2 to 6 weeks depending on complexity. Cost is calculated individually.

Work scope includes:

• Slot architecture documentation
• Source code with migrations
• Integration with your CRM or messenger
• Load testing (1000 RPS)
• Two-week post-launch support

Our Experience and Guarantees

We have implemented slot filling for 15+ projects in travel, fintech, and e-commerce. Experience with Rasa slot filling, Dialogflow slot filling, LLM fine-tuning, and other conversation AI tools — over 6 years. We guarantee full support during the implementation phase.

Contact us for a project assessment. Get a consultation today — we'll analyze your task and offer the optimal solution.

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.