Glossary

Below are plain-language definitions for common AI and regional terms to help non-technical readers.

• Token: A unit of text the model processes (roughly 3/4 of a word). Example: "banana" ≈ 1 token.
• Context window: The amount of text a model can see at once; longer windows allow longer prompts and documents.
• Hallucination: When a model invents facts or citations that are not true.
• Prompt injection: A malicious or accidental instruction embedded in input that changes the model's behavior.
• Fine-tuning: Training a model further on specific data so it behaves differently for a use-case.
• Instruction-following: How well a model follows explicit directions in a prompt.
• Latency: How long it takes the model to return a response.
• Throughput: How many requests a model can handle in a given time.
• Vector DB: A database that stores embeddings (numerical representations) for semantic search.
• Chain-of-thought: A prompting technique that asks the model to explain its reasoning steps.
• Provenance: Record of where information or data came from; important for verification.
• Explainability: How easy it is to understand why a model produced an output.
• Bias: Systematic favoring or disadvantaging of certain groups in outputs.
• AIFS: ASEAN Integrated Food Security framework (regional policy context).
• APTERR: ASEAN Plus Three Emergency Rice Reserve (regional food security mechanism).
• ASEAN: Association of Southeast Asian Nations; regional political and economic grouping.
• Claude / GPT / Llama: Examples of large language models (LLMs); use-case suitability varies.
• Tokens cost: Models often charge per token processed; keep prompts concise to control cost.
• Output confidence: A model's internal estimate is not always reliable; human verification is required.
• Hallucination mitigation: Steps to reduce invented facts: cite sources, cross-check, constrain outputs.

Add new terms as the site evolves. Mark any unverified definitions with [UNVERIFIED].