Griot
A team knowledge platform where every AI answer arrives with its receipts attached.
Overview
Griot is a team knowledge platform: upload the documents your company actually runs on, ask questions in plain language, and get answers where every claim carries a citation chip that opens the exact source passage, highlighted, in a reading drawer.
It's named for the West African griot — the keeper of a community's history whose authority rests on verifiable lineage. An answer without a source is just a rumor with good formatting.
The Challenge
Teams lose hours a week to searching their own documents, and the obvious fix — an AI that answers from them — collides with an operational reality: an unsourced answer can't be trusted with anything that matters. One hallucinated policy detail in a payroll or compliance question doesn't just fail; it poisons the team's trust in every future answer. The product's real adversary wasn't search latency — it was the first wrong answer.
Goals
- 01
Answer questions from team documents with every claim traceable to its source passage.
- 02
Make verification effortless: claim → receipt in one tap, with the passage highlighted in context.
- 03
Keep the workspace legible — documents, suggested questions, and answers in one calm surface.
- 04
Deliver the full ask-verify loop on mobile, where quick operational questions actually happen.
Research
The research examined how teams verify information when the stakes are real: they don't want summaries, they want the paragraph — the clause in the policy, the line in the contract. Verification behavior is about locating, not re-reading. That meant citations couldn't link to documents; they had to link to passages, scrolled-to and highlighted.
Studying RAG-product failures sharpened this: tools that cite at document level make verification a second search, and users stop verifying — which is exactly when hallucinations become dangerous. The citation had to be cheaper than trusting blindly.
Strategy Thesis
“Citations are the product — the answer is just the index into them.”
Griot inverts the usual RAG hierarchy. Instead of an answer with optional footnotes, it treats every response as a set of claims, each wearing its citation chip inline. Tapping a chip slides open a reading drawer with the source document scrolled to the passage, amber-highlighted — verification costs one tap and zero context.
The griot framing isn't decoration; it's the design principle. Authority through lineage: the product never asks to be trusted, it makes checking so cheap that trust accumulates as a side effect.
Design Process
- Step 01
Design the claim-chip grammar
Answers were designed as citable units from the start — the generation pipeline attaches source passages per claim, and the UI renders chips inline rather than as footnotes, so provenance is part of the sentence.
- Step 02
Make the receipt instant
The reading drawer opens over the conversation, source scrolled and highlighted, and closes back to the exact answer position — verification as a glance, not a navigation.
- Step 03
Seed the cold start
Workspaces generate suggested questions from the uploaded corpus, so a new team's first interaction is a one-tap demonstration of the full ask → answer → verify loop.
Final Solution & Gallery
The live product is a team workspace — document grid, search bar, suggested questions — where any question returns a cited answer, and the mobile flow keeps the whole loop intact: ask, read, tap a chip, see the passage.
Results & Metrics
The build's structural guarantees:
- of answer claims carry citation chips
- 100%
- citations open the source highlighted, not just the file
- Passage-level
- from any claim to its receipt and back
- 1 tap
- uncited assertions rendered as fact
- 0
Key Takeaways
- 1
Trust in AI answers is an interaction-cost problem: make verification cheaper than doubt and it actually happens.
- 2
Passage-level beats document-level citation by exactly the margin users care about.
- 3
A strong cultural metaphor (the griot's lineage) can discipline product decisions, not just name them.
Concept client, real build — every interaction shown is live in the linked product.