Glossary
AI terms, honestly defined
Every entry says what the term means and, where it applies, what people commonly get wrong about it. That second part is usually the more useful half.
- Terms
- 50
These definitions are written from building the things, not from summarising other glossaries. Where a term is routinely misused in vendor material, the entry says so plainly, because in practice the misunderstanding costs more than the ignorance does.
A worked example: a team that knows what RAG stands for but has never measured retrieval separately from generation will spend a quarter tuning prompts to fix an accuracy problem that lives entirely in the retrieval step. Another: a team that treats prompt injection as theoretical, right up until an agent with write permissions reads an email. The definitions below are written to head off exactly that class of expensive misreading.
If you are evaluating vendors, these entries double as questions worth asking. Anyone can define hallucination; far fewer can tell you their measured rate on your domain, or what their system does when confidence falls below threshold.
The misunderstandings that cost the most
Drawn from the caveat on each entry, the part vendors tend to leave out.
- AI agent
- Most products marketed as agents are workflows with a language model in one step. A real agent decides its own sequence; a workflow has the sequence decided for it.
- AI governance
- Governance written after deployment is documentation. Only governance designed alongside the system functions as a control.
- Chunking
- Chunking is usually treated as a configuration value. On real corpora it is one of the highest-leverage decisions in the whole pipeline.
- Computer vision
- A single headline accuracy number usually hides wide variation between defect classes. Per-class reporting is the honest form.
- Context window
- A large context window is often treated as a replacement for retrieval. It is usually slower, more expensive and less accurate than retrieving the right few passages.
- Evaluation
- Evaluation is the most commonly skipped step and the one whose absence is felt most, usually about three months in.
- Fine-tuning
- Most teams asking for fine-tuning need better prompting and retrieval instead. Fine-tuning genuinely wins on cost at volume, not usually on capability.
- Hallucination
- Teams frequently report that a model 'mostly gets it right' without having measured. The measured rate is usually higher than the impression.
- Human in the loop
- A rubber-stamp approval step is worse than none. It manufactures the appearance of control without the substance, which is precisely what an auditor looks for.
- Hybrid search
- Semantic search is often deployed as a replacement for keyword search, which reliably makes exact-match queries worse.
All terms
- AI agentautonomous agent, agentic AI
- AI gateway
- AI governance
- Chunking
- Computer vision
- Context window
- Data leakage
- Data residency
- DPDP Act
- Edge inference
- Embedding
- Evaluationevals
- Feature engineering
- Fine-tuning
- Guardrails
- Hallucination
- Human in the loop
- Hybrid search
- Inference cost optimisation
- Intelligent document processingIDP
- LoRAlow-rank adaptation
- MLOps
- Model Context ProtocolMCP
- Model drift
- Model routing
- Multi-agent system
- Multi-tenancy
- Object detection
- Observability
- OCR
- Open-weight models
- Predictive analytics
- Prompt engineering
- Prompt injection
- Quantisation
- RAGretrieval-augmented generation
- Red teaming
- Reranking
- Robotic process automationRPA
- Semantic caching
- Speaker diarisation
- Speech recognitionASR, speech to text
- Straight-through processing
- Structured output
- Text to speechTTS, speech synthesis
- Token
- Tool callingfunction calling
- Vector database
- Voice AI agent
- Workflow automation
Have a term we have not covered?
Tell us what you are trying to understand and we will give you a straight answer.
Or email bd@dtrasglobal.com · call +91 74118 77878