AIAdvanced#Infrastructure#Data Provenance#AI Safety
RAG (Retrieval-Augmented Generation)
R
Definition
A system pattern that retrieves selected external records and supplies them with a query to a language model. Retrieval can improve traceability and recency when sources are cited; it does not prove that a record is true, complete, current, or safe to act on.
Source reviewed
- NIST: Adversarial Machine Learning Taxonomy opens in a new tab
Supports: NIST identifies retrieval-augmented generation as a model-based system relevant to generative-AI attack classes, including indirect prompt injection and data poisoning.
- OWASP: LLM01 Prompt Injection opens in a new tab
Supports: OWASP documents that crafted untrusted content can alter an LLM application's behavior and recommends keeping application functions in code rather than granting model output direct authority.
Concept Diagram
Key Takeaways
1
RAG retrieves selected external records before a model drafts an answer; it does not guarantee live, correct, or complete information.
2
Show the original record, publisher, source and retrieval times, identifiers, selected excerpt, and retrieval rule beside the model summary.
3
Retrieved sources are untrusted input and must not change system instructions, permissions, or transaction policy.
4
Use RAG for reviewable research, not trading execution: high-impact actions require independent controls and explicit approval.
Practical Example
A user asks about a protocol upgrade. A RAG workflow returns the official notice and two reputable but conflicting reports, with the URL, publication time, retrieval time, and excerpts for each. It labels the conflict, drafts questions for review, and creates neither a trade alert nor a transaction.
Finished learning?
Related Terms
AI
Perception LayerThe part of an AI workflow that collects and labels inputs such as blockchain records, market data, and public communications. It supplies context for review; it does not establish truth, identify people, predict prices, or justify an action.
AI
LLM ReasoningUsing a language model to summarize, classify, or compare supplied information. Its output is probabilistic text, not verified market analysis, financial advice, or authority to execute a trade.
AI
Context WindowThe maximum input and output token capacity a model can handle in one request, subject to the model and provider configuration. It is a capacity limit, not dependable memory, data verification, or a measure of reasoning quality.
AI
Prompt EngineeringThe practice of designing model instructions, examples, context, and output constraints for a bounded task. It can make a workflow easier to evaluate; it does not make model output deterministic, correct, secure, or suitable for a financial decision.
AI
Vector DatabaseA system that stores vector representations and retrieves nearby vectors using a stated similarity measure. It can support semantic retrieval, but it does not determine whether retrieved content is true, current, or relevant enough to use.
Related Deep Dives
Master the hub of AI
Explore all our strategic guides about AI to take your operations to the next level.
View all articles