Document assistants
Q&A on policies and manuals.
What is a large language model (LLM)?
A large language model (LLM) is a language model trained at very large scale—billions of parameters, massive corpora—able to perform varied tasks few-shot: writing, assisted reasoning, code, dialogue. Work such as GPT-3 (Brown et al., 2020) showed these emergent abilities. Power comes with risks: hallucinations, prompt-based leaks, cost, and governance (NIST AI RMF, OWASP LLM).
In one sentence
An LLM is a large-scale language model that is versatile in few-shot settings.
Key points
- Scale and data broaden the task spectrum without systematic fine-tuning.
- Few-shot / instruction-tuning make the model “programmable” in natural language.
- It remains statistical: it can invent convincing facts.
- Production means prompts, RAG, evals, access controls, and token budgets.
Term at a glance
- Large language model
- LLM · Foundation language model
- English term
- Large Language Model (LLM)
- Domain
- Artificial intelligence
- Category
- Generative AI
- Level
- Intermediate
What does “large language model” mean exactly?
“Large” refers to scale (parameters, training tokens, compute). Transformers (Vaswani et al.) underpin most modern LLMs.
An LLM is not magically up to date: knowledge cutoffs and missing private data require RAG or tools.
For an SME, an LLM is an engine to frame—not an autonomous employee. Profitable cases: drafts, extraction, Q&A over validated internal docs.
How do you industrialize an LLM?
- 01
Frame the use case
Value, risk, need for cited sources.
- 02
Choose model and hosting
Cloud API vs private model by sensitivity.
- 03
Build the business layer
System prompt, RAG, tools, refusal policies.
- 04
Evaluate continuously
Golden sets, user feedback, cost and latency.
A concrete LLM example
An engineering firm in Montreal connects an LLM to an internal standards corpus (RAG). Engineers ask questions; answers cite excerpts. A committee reviews monthly errors. Document search time drops without letting the model invent a clause absent from the corpus.
What are LLMs used for?
Writing productivity
Drafts, rewrites, summaries.
Structured extraction
JSON from quotes, contracts, tickets.
Tooled agents
Supervised API calls (booking, CRM).
Benefits and limits of LLMs
- Few-shot versatility
- Fast prototyping
- High writing quality
- Tooling ecosystem (RAG, evals)
- Hallucinations
- Variable cost and latency
- Security risks (injection, exfiltration)
- Governance and compliance to build
How does an LLM differ from a classic ML model?
| Large language model | Classic supervised ML model | |
|---|---|---|
| Input | Free text, instructions | Tabular features / typed tensors |
| Output | Text / code / free JSON | Class, score, regression |
| Adaptation | Prompt, RAG, fine-tune | Retrain on labels |
| Typical risk | Fluent fabrication | Overfit / feature drift |
Why LLMs matter for Quebec SMEs
They compress time on language tasks—if you invest in sources, evaluations, and accountability. Otherwise you buy an expensive, risky demo.
Frequently asked questions
Open-source LLM or API?
API for speed; private if data is highly sensitive or residency is required. Hybrid is common.
Do we need to fine-tune?
Often not at first. RAG + strong prompts beat a rushed fine-tune.
How do we cite sources?
RAG architecture with retrieved passages and visible references.
Who owns errors?
The organization that deploys—hence policies, logs, and human review on critical cases.
Related terms
Sources and references
Considering an LLM assistant over your documents? We design RAG, evaluations, and guardrails for your risk level.
Industrialize your LLM