Understand
Retrieved Source Attribution
What sources did the system draw from?
Shows the documents the system looked up for an answer and which sentences they support, so people can check claims against the originals.
Overview
Why it matters
The sources exist. The interface just doesn’t show them.
A knowledge assistant gives a convincing answer, someone acts on it, and later nobody can say where it came from. Systems that look things up before answering do have sources. Without attribution design, people can’t check them.
This is about the documents the system looked up when it wrote this answer. For training and input data, see Data Provenance & Lineage.
What does it look like?
Check a specific claim
Check a specific claim
- Resident asks the benefits assistant about housing support and gets an answer with inline citation markers.
- Resident clicks [2] next to the sentence about required documents.
- Source panel expands showing the passage from “Housing support rules, city website, updated March 2026”, with match level “close”.
- Resident opens the original page on the city website.
Assess the sources behind an answer
Assess the sources behind an answer
- Resident sees “Based on 3 sources” above the answer.
- Resident opens the source panel.
- Resident sees one close match and two loose ones.
- Resident chooses “Talk to a person instead” to confirm the answer with a caseworker.
No source found
No source found
- Resident asks a question the city’s published benefit pages don’t cover.
- Assistant answers but shows “No source was found for this answer. It is based on the model’s general training and may be wrong. Check it, or ask a person.”
- Resident is offered a handover to a person.
When does it fit?
Use it when
- The system looks up documents before answering
- People need to check claims or cite sources
- Accuracy matters, as in legal, medical or policy work
Skip it when
- There is no lookup step
- Not even the type of source may be disclosed
- The task is casual and sources would only add noise
Design
How do I apply it?
-
Layer the sources.
A source count first, titles on request, full passages after that.
-
Tie citations to claims.
Markers next to sentences let people check each claim. Make sure each marker really supports its sentence.
-
Show the “no source” state.
If no source was found for an answer, say so.
-
Show how strong the match was.
A close match and a loose one should look different.
-
Sources aren’t proof.
A real source can still be misquoted. Pair with Grounding & Hallucination Indicators.
For your technical team
In retrieval-augmented generation, retrieval relevance (for example similarity scores) and answer faithfulness are separate measures. Surface relevance per source, and evaluate faithfulness separately before claiming an answer is supported. Similarity scores are not probabilities and aren’t reliably comparable across queries and can’t be compared across embedding models, so map them to two or three match levels (close, partial, loose) with thresholds set by the technical team.
How can I show it?
Source count summary
“Based on 3 sources” at a glance.
A compact indicator showing “Based on 3 sources”: how many sources were looked up for the answer. It does not show whether the answer reflects them. Pair it with grounding signals.
No source found state
Says plainly when no source was found.
An explicit design state for when no supporting material was found for an answer: “No source was found for this answer. It is based on the model’s general training and may be wrong. Check it, or ask a person.” This is not a failure state: it’s a transparency signal. Can apply to a single sentence as well as a whole answer. For public-service assistants, consider not answering and offering a person instead.
Inline citation markers
Markers next to the sentences they support.
Superscript numbers or footnote markers anchored to specific claims in the output, not just a list at the bottom. Users can see which sentence is tied to which source.
Relevance indicator
How closely each source matched the question.
A visual signal (dots or a short label) with two or three match levels: close, partial, loose. Helps users assess whether the retrieval was a strong match or a loose approximation. Avoid percentages, which read as “chance this is correct”.
Passage highlight
The exact passage, shown in context.
When a user clicks or hovers a citation marker, the retrieved passage is highlighted or shown in context, not just the document title.
Further left for everyday answers, further right for research and compliance work.
What does it cost?
Transparency vs. overload
Showing every passage by default can bury the answer. People may then skip the sources altogether. Show the count first and the passages on request.
Attribution vs. false confidence
A tidy source list can make an answer look more solid than it is. A source can be real and still not support the sentence it is attached to. Check that each marker supports its sentence, and pair sources with grounding indicators.
Disclosure vs. security
Showing documents can expose internal stores or restricted material. Even a title can reveal more than intended. Show source types or categories where the documents themselves can’t be shown.
Connections
Which patterns work with it?
Data Provenance & Lineage
Where does this data come from?
Data Provenance covers training and input data. This pattern covers documents looked up when the answer is written.
Grounding & Hallucination Indicators
Can I trust what the system generated?
Attribution shows what was retrieved; Grounding signals whether the output faithfully reflects it. Use together in high-stakes contexts.
Confidence & Uncertainty
How certain is the system?
Retrieval match levels complement output confidence signals.
Inspection Dialogue
Can I ask the system follow-up questions?
A user asking “where did you get that?” is a natural continuation of this pattern.
Sources (2)
Sources (2)
Lewis et al. (2020) — Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
coined the term RAG and showed that combining retrieval with generation improves knowledge-intensive tasks. Background for why runtime sources exist to be shown
Es et al. (2023) — RAGAS: Automated Evaluation of Retrieval Augmented Generation
proposes automated metrics for RAG pipelines, including faithfulness and context relevance. Useful for checking a pipeline before showing users any support signal
Your knowledge assistant answers, and nobody can say where the answer came from.
Internal search, policy assistants, support chatbots. That’s where this pattern starts. Let’s look at your interface together.
Work with me