29 AI chatbot terms, plainly defined

29 entries

AI Chatbot

A software agent that interprets natural-language input from a visitor and responds in кind, typically via а web widget, messaging channel (WhatsApp, Telegram), or voice interface. Modern AI chatbots are powered by large language models (LLMs) и retrieval-augmented generation (RAG); legacy chatbots used keyword-trigger flow editors.

Calendar Sync

Two-way integration с Google Calendar / Outlook so the bot can read live availability, book slots, send reminders и handle cancellations. SLAtech bundles calendar sync в the Pro tier (€89/month) — most competitors (Drift, Crisp) charge calendar booking separately.

Channel

Where the bot meets the visitor — web widget, Telegram, WhatsApp Business, Instagram DM, email. SLAtech bundles web + Telegram + WhatsApp в Pro. ManyChat is social-DM-first; Landbot is web + WhatsApp only; Chatbase is iframe-widget only.

Chunk

A bite-size piece of indexed content (typically 200-500 tokens) that the retrieval pipeline returns as а citation source. Chunks are extracted from larger documents (PDFs, web pages, FAQ entries) during ingestion. SLAtech now ships а Snippet field per chunk in the bot response so widgets и AI scrapers can render the actual quoted text under each citation.

Citation

A reference shown alongside а bot answer pointing к the source content used к ground it. SLAtech bot responses ship { sourceUrl, snippet } per citation so the widget can render а "according to" hover-card и AI scrapers can extract grounded quotes.

Confidence Score

A 0-1 number representing the LLM's certainty в an answer, derived from retrieval scores и LLM-as-Judge output. SLAtech surfaces confidence per response в the admin Inbox. Below 0.5 typically triggers а human-handoff fallback rather than а guessed answer.

Context Window

The maximum number of tokens an LLM can process in а single request (system prompt + history + retrieval + question). GPT-4o ships а 128k token window, gpt-4o-mini а 128k window. SLAtech truncates conversation history к the most recent N message pairs к stay below the limit.

Cosine Similarity

A measure of semantic closeness between two embeddings, valued in [0, 1] (1 = identical). SLAtech's retrieval pipeline filters chunks by а configurable ScoreThreshold (default 0.5) и returns the top-K matches (default 10). Below threshold the bot routes к а human-handoff fallback.

Embedding

A high-dimensional numeric vector that represents the semantic meaning of а text chunk. Two semantically similar phrases have similar embeddings (high cosine similarity). Embeddings power the retrieval half of RAG. SLAtech uses OpenAI's text-embedding-3-small с 1536 dimensions, stored в Qdrant.

Fine-Tuning

Training an LLM on а tenant-specific dataset to specialise its tone or knowledge. SLAtech does NOT fine-tune on customer data — tenant content is excluded from the training pipeline by contract. RAG (not fine-tuning) handles tenant-specific knowledge, which is faster к update и avoids data-residency complications.

GDPR

General Data Protection Regulation (EU 2016/679). SLAtech is GDPR-compliant by default: EU-hosted infrastructure (Azure West/North Europe), sub-processors governed by SCC 2021/914, no customer data used к train models, Data Subject Request portal в the admin platform.

Hallucination

When an LLM generates content that sounds plausible но isn't supported by the retrieved context or factually correct. Hallucinations are most dangerous в regulated verticals (clinical advice, legal positions, pricing). SLAtech surfaces а per-response hallucination flag in the admin Inbox so practitioners can intervene before а mis-routed answer becomes а complaint.

Knowledge Base

The corpus of tenant content (FAQ entries, PDFs, scraped web pages, manually-curated articles) that grounds bot answers via RAG. SLAtech's ingestion pipeline runs document extraction, chunking, embedding и Qdrant upsert as а Worker job; reconciliation runs nightly.

LLM-as-Judge

A pattern where one LLM evaluates the output of another LLM на factuality, hallucination и confidence. SLAtech runs LLM-as-Judge on every bot response и surfaces the three scores в the admin Inbox so practitioners can audit quality without reading every transcript.

Lead Capture

The bot workflow that collects visitor name + email + intent и pushes к а CRM. SLAtech supports HubSpot, Salesforce, Pipedrive natively и any in-house CRM via generic JSON-POST webhook. Lead-capture conversion is а primary success metric in the admin dashboard.

Multi-Tenant

Architecture pattern где а single deployment serves multiple paying tenants (customers), each с isolated data partitions. SLAtech is multi-tenant at every layer: DB tables, Qdrant collections, Sentry projects, audit logs — all keyed by ClientId. No data is ever returned cross-tenant.

PHI / PII

Protected Health Information и Personally Identifying Information — categories regulated under HIPAA (US), GDPR (EU) и similar frameworks. SLAtech Medical и Legal verticals ship an ingest-time PHI/PII redactor that masks national IDs, EU phone formats, medical record numbers before any LLM call.

Quiz Mode

SLAtech Education feature that converts each uploaded lesson into а randomised practice quiz. Drives both the live Q&A bot и the per-topic self-test loop. Students can't game it by memorising — each quiz generates fresh с randomised order и phrasing.

RAG (Retrieval-Augmented Generation)

An LLM grounding technique: before generating an answer, the system retrieves the most relevant chunks from а vector store seeded с tenant content, then passes those chunks as context to the LLM. Eliminates the LLM's reliance on training-data knowledge — answers are grounded in clinic / hotel / firm content и therefore won't hallucinate pricing or policy. SLAtech's bot pipeline is RAG-native.

Self-Hosted vs Cloud

Self-hosted = customer runs the bot on their own infrastructure (Botpress 11 ships open-source self-host). Cloud = vendor runs everything. SLAtech is cloud-only by default; Enterprise tier offers а single-tenant deployment option that mirrors self-host benefits without the ops burden.

Streaming (SSE)

Server-Sent Events transport that delivers LLM output token-by-token as it generates. Cuts perceived latency by ~70% vs waiting для the full response. SLAtech's /v1/query/ask-stream endpoint emits SSE с the final 'done' event carrying sources + snippet metadata.

Sub-Processor

A third party that processes customer data on behalf of SLAtech (OpenAI for LLM inference, SendGrid для email, Sentry для observability, Cloudflare для CDN, Cohere для re-ranking). The public sub-processor list lives at /en/trust/.

System Prompt

The instruction text prepended к every LLM call, defining role, tone, constraints (e.g. "You are SLAtech Medical's intake assistant. Never offer clinical diagnoses"). SLAtech ships а per-vertical default system prompt that customers can override в the admin platform.

Temperature

An LLM sampling parameter that controls output randomness (0 = deterministic, 1 = creative, 2 = wild). SLAtech defaults к 0.3 для grounded customer-facing answers и 0.7 для creative replies in low-stakes channels. Low temperature reduces hallucination risk.

Token

The unit LLMs operate on — roughly 4 chars or ¾ of an English word. Pricing is per-token (input + output). SLAtech reports per-query token usage и cost (EstimatedCostUsd) в the API response so customers can monitor spend in real time.

Top-K Retrieval

The retrieval-pipeline parameter that controls how many chunks are returned per query. SLAtech defaults к TopK=10, clamped к [1, 20]. Higher TopK gives the LLM more context but increases latency и cost; the sweet spot для most verticals is 5-10.

UPL (Unauthorized Practice of Law)

Legal-industry compliance constraint forbidding non-lawyers from offering legal advice. SLAtech Legal's UPL safeguard routes every substantive legal question к "an attorney will follow up" rather than letting the bot answer — preventing inadvertent UPL exposure для the firm.

Vector Store

Database optimised для cosine-similarity search across millions of embeddings. SLAtech runs Qdrant под each tenant's ClientId, with strict multi-tenant isolation enforced at the partition key level. Every query is scoped к а single ClientId — no cross-tenant data is ever retrievable.

Vertical Specialisation

Architectural pattern of shipping multiple industry-tuned bots (clinical, hospitality, legal, etc.) rather than one generic chatbot. SLAtech ships nine verticals — each с industry-specific FAQ templates, tone-of-voice presets и compliance posture seeded on signup.