AI is now in every vendor pitch and boardroom slide, yet terms get tossed around with more confidence than clarity. If you run a small firm or oversee operations, that gap creates risk: unclear promises, weak controls, and projects that drift. This guide gives you the one glossary you’ll need this year—and a practical way to use it in vendor conversations, pilot planning, and staff training. Wherever possible, we cite independent standards and research so you can speak precisely and make better decisions.
- How to use this glossary in vendor talks and internal planning
- Core model concepts: the words that explain how AI works
- Your data and retrieval: turning documents into answers
- Prompting and adaptation: getting outputs you can trust
- Safety, privacy and governance: the non-negotiables
- Side-by-side glossary cheat sheet
- Put the glossary to work: a 7-step vendor and rollout checklist
- Frequently asked questions
- Conclusion
How to use this glossary in vendor talks and internal planning
Use these terms as a script for due diligence. Ask vendors to show, not tell: “Which retrieval method do you use?” “How do you mitigate hallucinations?” “What’s your approach to PII?” Map answers to trusted frameworks such as the National Institute of Standards and Technology’s AI Risk Management Framework (AI RMF), which promotes shared terminology for transparency, explainability and governance in AI systems (nist.gov; airc.nist.gov; nist.gov).
If you’re screening suppliers, pair this glossary with your buying criteria. For a deeper vendor checklist tailored to small firms, see Selecting an A.I. Vendor.
Core model concepts: the words that explain how AI works
Transformer
The modern backbone of language and multimodal models. It relies on self-attention rather than recurrence to model relationships across tokens, introduced in the research paper “Attention Is All You Need” (arxiv.org; background summaries at wikipedia.org).
Parameter
A learned weight inside a neural network that adjusts during training; collectively, parameters encode the model’s knowledge. See foundational treatment in the free “Deep Learning” textbook by Goodfellow, Bengio and Courville (deeplearningbook.org).
Token and tokenization
Models process text as tokens—small pieces such as words or subwords—produced by tokenization algorithms like Byte Pair Encoding and SentencePiece (arxiv.org; arxiv.org).
Embedding
A numeric vector that represents meaning so that similar texts are close in vector space. Early word embeddings were popularized by word2vec (arxiv.org), with sentence-level embeddings advanced by Sentence-BERT (arxiv.org).
Context window
The maximum number of tokens a model can consider at once. Once your prompt plus prior conversation exceed this window, the model “forgets” earlier text (hai.stanford.edu). Vendors now advertise very long windows, but the concept remains the same: bounded working memory (ai.google.dev).
Temperature
A parameter that controls output randomness by adjusting the softmax distribution over next-token choices; higher temperatures produce more varied outputs. See Hinton et al. on distillation and temperature-scaled soft targets (arxiv.org) and textbook discussion of softmax behavior (deeplearningbook.org).
Your data and retrieval: turning documents into answers
Retrieval-augmented generation (RAG)
A technique that retrieves relevant passages from a knowledge base and feeds them to a generator model for grounded answers. Originally described in Lewis et al. (arxiv.org), RAG reduces reliance on the model’s internal memory by pairing it with your approved sources.
Vector database and ANN search
To support RAG at speed, embeddings are stored in indexes optimized for approximate nearest neighbor (ANN) search, such as FAISS and HNSW graphs. See Meta’s FAISS engineering note (engineering.fb.com) and the HNSW paper by Malkov and Yashunin (arxiv.org).
Chunking and citations
Documents are split into chunks that get embedded and retrieved; this improves recall and speeds search. Strong implementations return citations so users can verify claims—a practice encouraged by trustworthy AI guidance in the NIST AI RMF resource center (airc.nist.gov).
Prompting and adaptation: getting outputs you can trust
Prompt engineering and system instructions
Carefully crafted instructions and patterns that steer model behavior. Recent surveys catalog practical techniques and their trade-offs (arxiv.org; arxiv.org).
Zero-shot and few-shot
Getting a model to perform a task with no examples (zero-shot) or a small number of examples in the prompt (few-shot). The GPT-3 paper established few-shot performance as a widely observed capability (arxiv.org).
Fine-tuning
Updating some or all model parameters on labeled examples for your task. Fine-tuning improves domain adherence but demands careful evaluation and governance; see survey and alignment work for context (arxiv.org).
RLHF (reinforcement learning from human feedback)
Aligns models to human preferences by training a reward model on human judgments, then optimizing the generator against that reward. Foundational references include Christiano et al. on learning from human preferences and Ouyang et al. on instruction-following behavior (arxiv.org; arxiv.org).
Safety, privacy and governance: the non-negotiables
Hallucination
When a model produces ungrounded or false statements with high confidence. Multiple surveys document frequency, taxonomies and mitigation strategies across tasks (acm.org; arxiv.org).
Guardrails and application security
Defensive patterns—content filters, policy checks, tool and data-access constraints—that wrap the model. The OWASP Top 10 for LLM applications highlights critical risks like prompt injection and handling untrusted output (owasp.org; cheatsheetseries.owasp.org). For a deeper dive into operational exposure, see AI-Driven Cyberattacks Are Surging: What Small Businesses Must Do Now and Real-Time AI Security for Small Businesses and Law Firms.
PII (personally identifiable information)
Data that can identify an individual, directly or in combination. The formal guidance for U.S. organizations appears in NIST SP 800-122 (csrc.nist.gov).
Differential privacy
A mathematical guarantee that limits what can be learned about any one person from released data. Foundational references: Dwork’s original 2006 paper and the 2014 monograph by Dwork and Roth (harvard.edu; nowpublishers.com; see also NIST resources at nist.gov).
Zero trust
A security architecture that assumes no implicit trust, emphasizing continuous verification and least-privilege access. See NIST SP 800-207 and implementation guidance from the NIST National Cybersecurity Center of Excellence (csrc.nist.gov; pages.nist.gov).
Side-by-side glossary cheat sheet
| Term | Plain meaning | Where it helps in a small business |
|---|---|---|
| Transformer (source) | Neural network that models relationships across tokens with self-attention. | Modern assistants, drafting tools, code helpers. |
| Tokenization (source) | Splits text into pieces the model can process. | Budgeting usage, prompt design, data preprocessing. |
| Embedding (source) | Vector that captures semantic similarity. | Search across policies, emails, knowledge bases. |
| Context window (source) | The model’s working memory limit. | Knowing when to summarize, paginate, or use RAG. |
| Temperature (source) | Controls randomness in generation. | Marketing copy vs. policy language tone control. |
| RAG (source) | Retrieves facts, then generates. | Policy Q&A, product support, internal search. |
| Vector index (source) | Structure that finds similar embeddings quickly. | Fast knowledge lookup at scale. |
| Prompt engineering (source) | Systematic instruction patterns for better outputs. | Repeatable templates for briefs, emails, SOPs. |
| Few-shot (source) | Provide a handful of examples in the prompt. | Consistent formatting for invoices, summaries. |
| Fine-tuning (source) | Train on your labeled data for your task. | Specialized classification, style or compliance rules. |
| RLHF (source) | Learn preferences from human feedback. | Reduce unhelpful or off-policy responses. |
| Hallucination (source) | Confident but incorrect output. | Requires citations, RAG, and review steps. |
| PII (source) | Data that can identify a person. | Privacy reviews, redaction, retention controls. |
| Zero trust (source) | No implicit trust; verify continuously. | Access controls around AI tools and data. |
Put the glossary to work: a 7-step vendor and rollout checklist
- Define the job to be done. Write the outcome, source of truth, and failure modes. If automation is the goal, align to the concrete gains outlined in 7 Benefits of Process Automation for Small Business Owners.
- Choose the interaction pattern. For lookup-style tasks, prefer RAG to reduce hallucinations and ensure citations (arxiv.org).
- Set guardrails up front. Map risks to OWASP’s LLM Top 10 and implement prompt-injection defenses and output handling (owasp.org).
- Decide on adaptation. Try zero-/few-shot prompts first; consider fine-tuning only if accuracy plateaus (arxiv.org; arxiv.org).
- Plan for privacy. Identify PII in prompts, responses, and training data; apply redaction and, where appropriate, differential privacy methods (csrc.nist.gov; nowpublishers.com).
- Enforce access with zero trust. Restrict tools and data by role, log usage, and review permissions regularly (csrc.nist.gov).
- Pilot, measure, then scale. Track quality, latency, and user satisfaction. Where your firm already uses Microsoft 365, see how governance and AI features can be rolled out responsibly in Enable Claude in Microsoft 365 Copilot for Law Firms.
Frequently asked questions
What’s the difference between RAG and fine-tuning, and which should I try first?
RAG retrieves your approved content at query time, so answers can cite sources and update as your documents change. Fine-tuning changes model behavior using labeled examples. For small businesses, start with RAG for factual tasks, then consider fine-tuning if prompts and retrieval no longer improve accuracy.
How do I reduce hallucinations without hiring a full data science team?
Use retrieval-augmented generation for factual questions, return citations, and add a lightweight review step for high-stakes outputs. Implement basic guardrails for prompt injection and output handling following OWASP’s guidance. These steps address common failure modes documented in surveys of LLM hallucinations.
What is a context window in plain English, and why do vendors brag about it?
It’s the model’s working memory: how much text it can consider at once. Bigger windows can fit more background material, but quality still depends on retrieval quality and prompt design. Even large windows benefit from concise prompts and chunked, relevant context rather than everything at once.
Is collecting customer data for AI features a privacy risk for a small firm?
Yes, if you can reidentify individuals or keep more data than you need. Follow NIST SP 800-122 to classify PII, minimize collection, and set retention. If you share aggregate stats or train with user data, consider differential privacy techniques and a zero-trust approach to system access.
Conclusion
AI fluency is now a business skill, not a research hobby. With a shared glossary—rooted in standards and primary research—you can ask sharper questions, scope safer pilots, and buy with confidence. Start with the essentials: retrieval over raw recall, prompts before fine-tuning, and security as a first principle. Then measure what works and codify it into your operations. If you want help applying this to your workflows, we’re ready to partner.
Ready to explore how you can streamline your processes? Reach out to A.I. Solutions today for expert guidance and tailored strategies.
