Build a Research-Repository RAG That Cites or Refuses

Workflow area: Research & Synthesis · Tier: Core · Hands-on time: ~40 minutes

Every research team has the same graveyard: dozens of studies, decks and Dovetail projects that nobody re-reads, so the same questions get re-researched or — worse — answered from memory. Retrieval-augmented answers fix the access problem, but a naive setup creates a subtler one: an assistant that always answers, blending a 2022 study with a 2026 one and improvising when the corpus is silent. The design goal is an assistant whose most valuable output is sometimes "we don't know" — because a research repository that never refuses is just a hallucination engine with your logo on it.

When to use — and when not to

• Use it when your repository has critical massten or more studies with real findings; below that, a well-organized index page serves better than retrieval.
• Use it to arm non-researchersPMs and designers asking "what do we know about X?" at the moment of decision is the entire point; the citation trail keeps their confidence proportionate.
• Use it to expose research gapsevery logged refusal is a documented unknown; the refusal log is your next research roadmap, written by demand.
• Skip it if the corpus is stale or contestedretrieval amplifies whatever is in the corpus; garbage studies retrieved confidently are worse than no tool.
• Skip the DIY build if Dovetail's native Ask already covers youconfigure and audit the platform feature before wiring a custom pipeline; the discipline transfers either way.

The recipe

1. 1.

   Curate the corpus — this is the step that decides everything. Inventory your studies and admit only documents that state their method, date, and participant base. For each admitted study, create a summary sheet: study title, date, method, N, key findings as discrete numbered claims. Exclude drafts, brainstorms and synthetic-user output entirely — if it enters the corpus, it will someday be cited as evidence.

2. 2.

   Stand up retrieval where your repository lives. In Dovetail, that is enabling Ask over curated projects only. In Notion AI, a dedicated "Research Findings" teamspace the AI answers from. For a custom route, a Claude Project with the summary sheets as project knowledge works for corpora up to moderate size. Whatever the vehicle, the corpus boundary must be explicit: the assistant answers from admitted documents, nothing else.

3. 3.

   Install the answering contract as a system prompt. Adapt:

   Copy

   You answer questions about our user research using ONLY the studies
   in this corpus. Rules:
   
   1. CITE: every claim names its study and date, e.g. (Checkout study,
      Mar 2025, n=12). Quote participant evidence verbatim when it exists.
   2. WEIGHT RECENCY: when studies conflict, prefer newer evidence and
      say so: "The 2024 study found X, but the 2026 study found Y."
   3. REFUSE ON THIN EVIDENCE: if fewer than [2] studies address the
      question, or evidence is older than [18 months] for product-
      specific questions, say: "The repository doesn't have solid
      evidence on this. Closest related finding: [...]. This would be
      a new research question." Never fill gaps with general UX
      knowledge presented as our research.
   4. SCOPE HONESTY: state each cited study's method and N; a 6-person
      diary study and a 2,000-response survey do not carry equal weight.
   5. NEVER blend our findings with industry common sense without
      labeling which is which.
   

4. 4.

   Red-team it before announcing it. Ask ten questions: four you know the corpus answers, three it half-answers, three it cannot answer. Grade each response — citations real? Recency weighted? Refusals actually refused? A good refusal looks like:

   Copy

   Q: What do we know about how enterprise admins feel about SSO setup?
   
   A: The repository doesn't have solid evidence on this. Closest
   related finding: IT admins described initial configuration as
   "high-stakes, low-frequency" work (Admin console study, Jan 2025,
   n=8) — but that study did not cover SSO specifically. This would
   be a new research question.
   

   Then verify two citations per answered question against the source studies. A single fabricated citation means the contract needs tightening before launch, not after.

5. 5.

   Operate it: log refusals, refresh the corpus, re-audit quarterly. Route refusals to a "research gaps" list your team reviews when planning studies. When new studies land, admit them through the same curation gate. Once a quarter, rerun the ten-question red-team — retrieval quality drifts as corpora grow, and the audit habit is what keeps trust calibrated.

Make it yours

• Dovetail-native version: curate projects into an "evidence-grade" collection, enable Ask over only that collection, and pin the answering contract in your team's usage guide — the refusal rules live in how people are taught to read answers.
• Claude Project for the lean team: one Project, summary sheets as knowledge, the contract as instructions; costs nothing extra and the corpus boundary is exactly the file list.
• Stale-claim sweep tie-in: run pb-03's evidence-link discipline here too — each quarter, ask the assistant which cited findings are older than 18 months and route the list to your persona/journey-map refresh.

Quality check

• The corpus admission gate is written downwho admits studies, and on what criteria; an uncurated corpus fails silently and permanently.
• Spot-checked citations trace to real passagesa citation that names a real study but misstates its finding is the failure mode retrieval hides best.
• Refusal thresholds are explicit numbers"thin evidence" undefined means the model decides, and models hate refusing.
• Method and N ride along with every answerfindings stripped of their sample size become universal truths by the third re-telling.
• The refusal log feeds research planningif nobody reads the gaps list, the refusals were theater.