Founder Guide
RAG System Planning Questionnaire
A Retrieval-Augmented Generation (RAG) system lets an AI answer questions using your own content instead of guessing. Describe your use case, choose the subsystems your application needs, then work through each one. Switch to the technical perspective when you want the engineering questions to appear alongside the business decisions. When you are done, generate a build prompt to hand to an AI coding agent.
1. What is this for?
A short description of the product or workflow you are planning for. It appears at the top of your generated build prompt.
2. Choose your subsystems
These are the building blocks of a RAG system. Keep the ones your application needs and turn off the rest. Only the subsystems you keep show below and count toward your plan. Not sure? Leave it on and decide on the items inside.
3. Perspective
Technical adds engineering questions inside each section. They appear under an item once you include it.
1
Inputs: source types and formats
Choose the source types you want, then specify the exact formats and how each should be handled. Reading, extracting, and splitting each type into searchable pieces is part of supporting it.
Text documentsReports, guides, contracts, and notes.
Web contentPages from your site or the open web.
Spreadsheets and structured dataTables, exports, and data files.
ImagesPhotos, screenshots, diagrams, scans, and charts.
AudioCalls, interviews, voice notes, and podcasts.
VideoRecordings, demos, and webinars.
Connected business appsLive data from tools your team already uses.
Email and chat historyThreads from inboxes and messaging tools.
Code and technical docsSource code and developer references.
Content in other languagesMaterial that is not in English.
2
Ingestion and freshness: how content arrives and stays current
Decide how material gets in and how you keep it from going stale.
Manual uploadA person adds files when needed.
Automatic syncNew or changed content flows in from connected sources on its own.
Scheduled refreshRe-checks sources on a set cadence.
Near real-time updatesChanges show up within seconds or minutes.
DeduplicationAvoids storing the same content twice.
Version historyKeeps older versions when a document changes.
Carry over source permissionsHonors who could see a file in its original place.
3
Enrichment: what gets added and kept current
Optional steps that add structure and meaning. For several, you can choose whether it is built once or kept maintained as content changes.
Auto-summariesA short summary generated for each document.
User taggingLets people add and edit their own tags.
System taggingTags the system applies automatically.
Metadata generationFills in missing details like title, date, or author.
Detecting and hiding sensitive infoFinds personal or confidential data and masks it.
Language detection and translationSpots the language and can translate when needed.
Freshness and quality scoringFlags outdated or low-quality material.
Taxonomy generation and maintenanceBuilds a category structure for your content and keeps it current as it grows.
Entity extraction and maintenanceIdentifies people, companies, products, dates, and amounts.
Relationship mapping and maintenanceTracks how those entities connect to each other.
4
Indexing and retrieval: what is searchable and how
Spell out what a query can actually reach, and the way the system finds it.
What can be searchedPick which parts of your content a query can reach. The choices reflect the enrichment you turned on.
Search methodHow matches are found.
FiltersLet people narrow results before or after searching.
Result depth controlHow much content to pull in for each answer.
Re-rankingA second pass that pushes the most relevant results to the top.
Separating data by customer or teamKeeps one group's content out of another's results.
5
Generation: how answers are produced
How the system turns retrieved content into a response.
Choice of AI model and providerWhich underlying model writes the answers, and who runs it.
Always cite sourcesEvery answer points back to the documents it used.
Admit when the answer is not thereSays it does not know instead of guessing.
Tone and persona controlSets how formal, brief, or on-brand the voice is.
Conversation memoryRemembers earlier turns so follow-up questions make sense.
Confidence signalsShows how sure the system is about an answer.
6
Outputs: what users get back
The forms an answer can take. Pick everything you want to offer.
Smart answersA direct, written answer to the question.
Citations and referencesLinks back to the exact sources used.
Text summariesCondensed overviews of one document or many.
Suggested follow-up questionsPrompts that help people go deeper.
Raw file accessLets people open or download the original file.
Wiki viewA browsable, linked encyclopedia of your content.
Knowledge graph viewA visual map of topics and how they connect.
Audio overviewA generated spoken briefing on the material, similar to NotebookLM.
Video overviewA generated video walkthrough of the material.
7
Delivery: where it shows up
The places people and other systems reach it.
Chat interfaceA conversational window people type into.
Search boxA single box that returns answers and sources.
Bot in a chat platformAnswers questions inside your team or customer chat tools.
Embedded in your productBuilt into a screen your users already use.
Feedback captureThumbs up or down, so you learn what is working.
API accessLets other software send questions and get answers.
MCP serverExposes the system over MCP so AI assistants and agents can query it directly.
8
Access and permissions: who can see what
Set the access levels and the structure that enforces them. This keeps the right content in front of the right people.
Identity and sign-inHow people prove who they are.
User rolesTypical roles that bundle a set of rights.
Access tiersSensitivity levels you assign to content.
How access is grantedThe model that decides who gets in.
Permission granularityThe smallest thing you can control access to.
Enforce access at search timeRestricted content never appears in results for people who lack rights.
Guest or external accessLimited access for people outside your organization.
Redaction by clearanceThe same document can show more or less depending on who is viewing it.
Permission administrationWho can change access rights, plus a log of those changes.
9
Quality and evaluation: proving it works
How you check that answers are accurate, before and after launch.
Test question setReal questions with known good answers to measure against.
Grounding checksVerifies answers are backed by source content, not invented.
Human review loopPeople spot-check answers, especially early on.
Ongoing monitoringTracks answer quality over time, not just at launch.
Test changes before shippingCompares versions so an update does not quietly make things worse.
10
Operations, cost, and trust: running it responsibly
The practical and legal realities of keeping it live.
Hosting choiceA vendor service, your own cloud, or fully on your premises.
Security basicsEncryption and tight control over who has admin access.
Cost controlsA usage budget and caching to keep spend predictable.
Retention and deletionHow long content is kept and how it gets removed.
Data location and complianceWhere data lives and which rules apply.
Audit loggingA record of who asked what and what the system did.