Grounding
Grounding constrains or connects an AI output to relevant evidence, data, rules, or observed state available for the task.
Download Markdown (.md)Soroush’s Product Wiki · https://wiki.soroush.uk/topic/grounding/
01What it means
Grounding is the practice of giving a model authoritative context and requiring outputs to be supported by it. Evidence may come from retrieved documents, databases, tool results, user-provided material, or the current environment. Grounding differs from fine-tuning: it supplies task-time evidence rather than primarily changing model parameters.[1][2]
- 01User taskQuestion, identity, permissions, and current context
- 02Retrieve evidenceRelevant records from authorized, current sources
- 03Assemble contextSource boundaries, provenance, and instructions
- 04Generate responseUse evidence while preserving uncertainty
- 05Verify + citeCheck claims, permissions, conflicts, and missing support
- 06Monitor failuresImprove retrieval, sources, prompts, or workflow
Grounding reduces unsupported output only when retrieval, evidence quality, permissions, instructions, and verification work together; it does not guarantee truth.
02Why it matters
Grounding can make answers more current, domain-specific, traceable, and correctable. It also lets a product show users where a claim came from, but only if retrieval, source quality, claim support, and access control are designed together.
03When to use it
Use grounding when answers depend on changing, proprietary, or attributable information. It is less relevant for unconstrained creative work, and insufficient when the task requires guaranteed calculations, formal validation, or access decisions that deterministic systems should perform.
04How it works
- Identify authoritative sources, owners, freshness requirements, and user access boundaries.
- Select relevant evidence using retrieval, queries, tools, or explicit user input.
- Present evidence with clear provenance and instructions that distinguish it from untrusted commands.
- Require claim-level citations, abstention, or structured fields that can be checked.
- Evaluate retrieval coverage and whether generated claims are actually supported, then monitor both layers.
05A practical example
An internal policy assistant retrieves only documents the employee may open, passes dated excerpts to the model, and links each answer sentence to a source. If sources conflict or no passage supports the question, it says so and offers the policy owner's contact.
06Where it breaks
Grounding cannot repair missing, biased, stale, or incorrectly retrieved evidence. A model may ignore context or attach plausible but invalid citations. Retrieved content can contain prompt injection, and broad retrieval can leak information unless authorization is applied before content reaches the model.
Common mistakes
- Calling an answer grounded merely because documents were included in the prompt.
- Evaluating generation quality while ignoring retrieval failures.
- Retrieving first and filtering permissions after model generation.
- Assuming a displayed citation necessarily supports the nearby claim.
07Sources
- Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence ProfileNational Institute of Standards and Technology · Government frameworkArtificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile
- Retrieval-Augmented Generation for Knowledge-Intensive NLP TasksAdvances in Neural Information Processing Systems · Peer-reviewed researchRetrieval-Augmented Generation for Knowledge-Intensive NLP Tasks