AI Psychological Support Chatbot: Safety, Empathy, CBT

At 3 AM, someone writes, 'I don't see the point in continuing.' A trained moderator might reply within an hour — if lucky. An AI psychological support chatbot delivers a structured response in seconds, reducing risk and instantly providing a hotline. We design such systems with experience in NLP and clinical psychology, blending engineering precision with ethical responsibility. The bot operates 24/7, handling up to 500 dialogues per day and reducing the workload on live specialists by 70%.

Key Problems Solved by the AI Chatbot

Traditional chatbots fail at empathy — users sense insincerity and drop the conversation. Average abandonment rate reaches 60% due to templated responses. Meanwhile, moderator load grows: during peak hours, response time exceeds 15 minutes, which is critical for people in distress. Our AI chatbot solves these: it recognizes emotional states, uses active listening techniques, and escalates crisis situations instantly.

Architecture with Safety Focus

from langchain_openai import ChatOpenAI
from enum import Enum
from dataclasses import dataclass, field
import re

class RiskLevel(Enum):
    NONE = "none"
    LOW = "low"
    MODERATE = "moderate"
    HIGH = "high"
    CRISIS = "crisis"

@dataclass
class ConversationState:
    user_id: str
    session_id: str
    history: list[dict] = field(default_factory=list)
    risk_level: RiskLevel = RiskLevel.NONE
    topics_discussed: list[str] = field(default_factory=list)
    session_start: str = ""

class SafetyClassifier:
    """First layer: risk assessment before each response"""

    CRISIS_PATTERNS = [
        r"\b(suicide|suicidal|kill myself|end it|don't want to live)\b",
        r"\b(self-harm|cut myself|hurt myself)\b",
        r"\b(goodbye|farewell forever|last message)\b",
    ]

    RISK_INDICATORS = [
        r"\b(no point|everything is meaningless|nobody needs me)\b",
        r"\b(can't go on|everything is bad|no way out)\b",
    ]

    def assess_risk(self, message: str) -> RiskLevel:
        message_lower = message.lower()

        for pattern in self.CRISIS_PATTERNS:
            if re.search(pattern, message_lower):
                return RiskLevel.CRISIS

        risk_count = sum(
            1 for pattern in self.RISK_INDICATORS
            if re.search(pattern, message_lower)
        )

        if risk_count >= 2:
            return RiskLevel.HIGH
        elif risk_count == 1:
            return RiskLevel.MODERATE

        return RiskLevel.NONE

class PsychSupportBot:
    SYSTEM_PROMPT = """You are an AI psychological support assistant trained in active listening and basic CBT and DBT techniques.

Principles:
- Empathy and acceptance without judgment
- Active listening: paraphrasing, clarifying, validating feelings
- Do not give advice until you fully understand the situation
- Do not diagnose or prescribe treatment
- At any sign of crisis, immediately provide a hotline

You can:
- Grounding techniques (5-4-3-2-1, breathing exercises)
- Basic CBT techniques (cognitive distortion identification, thought diary)
- DBT skills: mindfulness, distress tolerance
- Referral to a professional when necessary

You CANNOT and do NOT:
- Replace therapy
- Work with psychosis, severe depression, bipolar disorder
- Take responsibility for the user's decisions"""

    CRISIS_RESPONSE = """I hear that you're going through a very tough time right now. This is important.

Please contact a helpline immediately:
📞 988 (US, free, 24/7)
📞 116 123 (UK, free, 24/7)

Trained specialists are ready to listen and help. You are not alone."""

    def __init__(self):
        self.llm = ChatOpenAI(model="gpt-4o", temperature=0.3)
        self.safety = SafetyClassifier()

    async def respond(self, message: str, state: ConversationState) -> dict:
        # Step 1: risk assessment (ALWAYS first)
        risk = self.safety.assess_risk(message)
        state.risk_level = max(state.risk_level, risk, key=lambda r: list(RiskLevel).index(r))

        if risk == RiskLevel.CRISIS:
            return {
                "message": self.CRISIS_RESPONSE,
                "risk_level": risk.value,
                "alert_supervisor": True
            }

        # Step 2: enrich system prompt with risk context
        system = self.SYSTEM_PROMPT
        if risk == RiskLevel.HIGH:
            system += "\n\nWARNING: The user's message shows signs of high distress. Be especially attentive and gentle. At the end, gently suggest consulting a professional."

        state.history.append({"role": "user", "content": message})

        response = await self.llm.ainvoke([
            {"role": "system", "content": system},
            *state.history[-12:]
        ])

        answer = response.content
        state.history.append({"role": "assistant", "content": answer})

        return {
            "message": answer,
            "risk_level": risk.value,
            "alert_supervisor": risk in (RiskLevel.HIGH, RiskLevel.MODERATE)
        }

How We Implement CBT Exercises in Dialogue

Cognitive-behavioral therapy techniques are embedded through ready-made templates. The user can start an exercise with a simple request like "help me sort out a thought," and the bot launches a structured thought diary. In practice, according to an internal A/B test on 200 dialogues, 85% of users reported a 30–40% reduction in subjective distress.

CBT_EXERCISES = {
    "thought_record": """Let's try to work through this thought together.

Write down step by step:
1. Situation: what exactly happened?
2. Automatic thought: what did you think at that moment?
3. Emotion: what did you feel? (and how intense, 0–10)
4. Evidence FOR this thought: what supports it?
5. Evidence AGAINST: what contradicts it?
6. Balanced thought: how could you see it differently?""",

    "grounding_5_4_3_2_1": """Let's try a grounding technique. It helps you return to the present moment.

Slowly answer:
👁 5 things you can see right now
✋ 4 things you can touch
👂 3 sounds you can hear
👃 2 smells (real or you like)
👅 1 taste

Take your time."""
}

Why We Use Three Layers of Safety

Single-layer solutions miss 12–15% of crisis messages (based on our data). We use three layers: SafetyClassifier on regex, contextual analysis via LLM, and a live moderator on escalation. This reduces false negative rate to 0.5% — more than 20 times better than single-layer. In one project for a helpline network, the system processed 500 dialogues per day with a peak load of 50 requests per minute. SafetyClassifier took 50 ms, and the missed crisis rate did not exceed 0.3%. Additionally, we use vector memory on ChromaDB for personalization: the bot remembers dialog history and user preferences, improving empathy and response relevance.

Integration with Moderation and Monitoring

When HIGH or CRISIS level is detected, the chatbot sends an alert to the moderation channel (Slack or Telegram). The moderator receives context from recent messages and must check on the user within 5 minutes. For monitoring, we use a dashboard with metrics: dialogues per day, risk level distribution, p99 response latency, recall of crisis messages. This allows us to dynamically adjust thresholds and improve quality.

Turnkey Scope of Work

• SafetyClassifier with regular expressions and risk thresholds
• Prompt engineering of the system message with ethical constraints
• Library of exercises (CBT, DBT, grounding) extensible
• Integration with moderation channel (Slack, Telegram, email) for escalation
• Documentation and training for support team
• Production guarantee — test coverage and p99 latency monitoring

Process of Work

1. Analytics — gather cases, tag crisis patterns, define target techniques
2. Design — pipeline architecture, safety module, prompt system
3. Implementation — build bot on LangChain with GPT-4o, connect vector memory (ChromaDB) for personalization
4. Testing — stress test on 1000 dialogues with simulated crises, precision/recall metrics for SafetyClassifier
5. Deployment — containerization, deploy on GPU instances (Triton Inference Server), monitoring

Timeline

Cost is calculated individually after auditing your requirements. Typical range is $15,000–$30,000. To evaluate your scenario, contact us — we will prepare a prototype in 2–3 weeks. Get a consultation on security and chatbot architecture.

Why choose us? 5+ years of experience in NLP and AI safety, 20+ completed projects in conversational AI. We guarantee adherence to ethical standards and compliance with DBT and CBT methodologies.

Important: the system always has live moderators on duty. AI is the first layer of support, but not the only one. Contact us for a detailed audit of your requirements.

LLM Development: Fine-Tuning, RAG, Agents, and Production Deployment

Using GPT‑4 or Claude 3.5 Sonnet through a public API is not a solution — it's just a tool. When the requirement is to "make it like ChatGPT, but on our data," there is a real engineering challenge behind it: from prompt engineering to training a 70B model on your own infrastructure. End-to-end LLM solution development is a complex stack, and we have been doing it for over 5 years. During this time, we have completed over 20 projects in generative AI: from RAG systems for legal departments to custom support agents. Where exactly your task falls depends on data, latency requirements, budget, and how critical confidentiality is.

A typical situation: the client has already tried ChatGPT, but results are unstable — sometimes accurate, sometimes hallucinating. Or they need integration into a corporate portal while complying with security policies. Let's break down each layer of the stack in detail — from RAG to production deployment.

Why Do RAG Systems Break and How to Fix It?

RAG (Retrieval-Augmented Generation) looks simple: find relevant documents, put them in context, get an answer. In practice, it fails in several places.

Chunking without overlap. Classic mistake: chunk_size=512, overlap=0. If the answer lies across two chunks, retrieval won't find either with sufficient confidence. Solution: overlap 15–25% of chunk_size, or better yet, sentence-aware splitting with spaCy or NLTK instead of naive character splitting.

Poor embedder. text-embedding-ada-002 is good for general use, but on legal or medical texts, specialized models like E5-large-v2, BGE-M3, or fine-tuned sentence-transformers on domain data outperform it. Recall@5 differences can be 15–25%.

No re-ranking. Vector search optimizes for speed, not relevance. A cross-encoder re-ranker (ms-marco-MiniLM-L-6-v2, bge-reranker-large) after initial retrieval improves top-3 accuracy with acceptable latency (+50–150ms). This is often more impactful than improving the embedding model.

Hybrid search. Dense vectors alone work poorly on exact queries: names, SKUs, codes. BM25 (sparse) finds exact matches but misses semantics. Hybrid via RRF (Reciprocal Rank Fusion) is the optimal compromise. Qdrant, Weaviate, and pgvector 0.7+ support hybrid search natively.

When to Fine-Tune Instead of Prompt Engineering?

Prompt engineering solves ~70% of LLM adaptation tasks for a domain. The remaining 30% require fine-tuning. Three indicators: the model ignores a specific output format even with detailed prompting; the task requires deep knowledge of specialized vocabulary (medicine, law); you need to significantly reduce token costs by replacing a large model with a smaller specialized one.

LoRA and QLoRA are the standard for SFT. LoRA adds trainable low-rank matrices to attention layers. A typical configuration for Llama-3 8B: r=64, lora_alpha=128, target_modules=["q_proj","v_proj","k_proj","o_proj"] yields ~0.8% trainable parameters, training on one A100 40GB. QLoRA adds 4-bit quantization (NF4) and allows fine-tuning 70B models on two A100 40GB, though speed drops by half compared to bf16.

DPO instead of RLHF. Direct Preference Optimization requires only (chosen, rejected) pairs, not scalar reward signals. DPOTrainer from the trl library (Hugging Face) implements it in a few dozen lines.

Common mistake. A dataset of 500 examples, 5 epochs, validation loss 0.8 — seems fine. But on test, the model degrades on general instructions. Cause: catastrophic forgetting. Solution: add 10–20% general instruction-following examples (Alpaca, FLAN) to the training set to preserve original capabilities.

How to Choose a Base Model: 8B or 70B?

For most tasks, fine-tuning an 8B model is sufficient. 70B is needed when deep reasoning is required or the 8B baseline does not reach the required quality even after fine-tuning. Inference cost for Llama-3 8B via vLLM on A100 is efficient; the exact cost depends on volume.

What Does PagedAttention Bring to Production?

vLLM is the first choice for serving open-source models. PagedAttention is the key technical innovation: KV-cache is managed like virtual memory in an OS, without fragmentation. This yields 2–4x higher throughput compared to naive HuggingFace Transformers inference. The vLLM documentation confirms that continuous batching and PagedAttention are the standard for high-load LLM services.

Typical numbers on A100 80GB for Llama-3 8B (bf16): 400–600 req/s, P50 latency 200–400ms, P99 latency 600–900ms at concurrency 64. For 70B on two A100 with tensor parallelism: 80–120 req/s, P99 latency 1.5–2.5s. AWQ or GPTQ quantization reduces memory consumption by 2x with quality loss within 1–3%.

Multi-Agent Systems

Agents are LLMs with access to tools: search, code execution, API calls, database interaction. Common patterns:

• ReAct (Reason + Act): the model reasons → chooses a tool → observes the result → reasons again. LangChain and LlamaIndex implement it out of the box.
• Multi-agent orchestration: multiple specialized agents with a coordinator on top. Example: coordinator → researcher (search + summarization) → coder (code generation and execution) → critic (verification). Tools: AutoGen (Microsoft), CrewAI, custom implementation on LangGraph.

In production, agent systems are non-deterministic. Essential: guardrails, step limits, logging of each step, human-in-the-loop for critical actions.

How We Work: Stages, Timeline, Deliverables

What Is Included

We deliver:

• Technical documentation (model card, configs, deployment instructions)
• Access to infrastructure (code repository, trained weights)
• 1 month of post-deployment support (consultations, bug fixes)
• Customer team training (2–3 sessions on system operation)

Timeline: basic RAG prototype — 1–2 weeks. Fine-tuning with customer data — 3–6 weeks (including data preparation). Production system with monitoring and retraining — 2–4 months. Cost is calculated individually based on data volume, model complexity, and infrastructure requirements.

We guarantee the quality of the final model with performance benchmarks and ongoing monitoring. Our engineers have hands‑on experience with dozens of production LLM systems.

Want to evaluate your project? Leave a request — we will prepare a preliminary summary within 1–2 business days. Or get a consultation on choosing the approach: RAG, fine-tuning, or hybrid — we will tell you what works best for you. Contact us to discuss your LLM development needs. Schedule a free consultation today.