Folio

Folio answers questions from your company's own documents and shows the exact passage behind every sentence, or says it has no source.

Made withRAGHybrid search (BM25 + vector)RerankingGolden-set evaluation

Watch the film Build something like this

Facts

Status: Live
Platform: Web app + mobile
Flows: Ask · Collections · Evaluations · Settings
Engine: Hybrid search (BM25 + vector) · reranker · citation threshold · answer model with numbered citations
Sources: PDF · DOCX · XLSX · web pages; OCR for scanned PDFs
Year: 2026

FIG. 01 / PRODUCT FILM

Muted preview · play with sound for the full cut

Who it's for

Teams whose answers live in their own documents and have to be backed by a source: an operations or HR lead fielding policy questions, legal counsel checking contract terms, a finance controller asked about travel rates. And the admin who keeps the collections indexed and tests every retrieval change before it reaches them.

The problem

Company knowledge sits in handbooks, policies, contracts and wikis. Ask a chat assistant about it and the answer arrives fluent and unchecked: nobody can see which page it came from, whether that document is still current, or whether the model filled a gap with a guess. And when someone changes how documents are searched, nothing shows whether the answers got better or worse.

The product

You choose the collections to search and ask in plain words. Folio runs a hybrid keyword and vector search, reranks the candidates and passes only the passages above a citation threshold to the answer model, which answers sentence by sentence, each with a numbered citation that opens its source page with the cited words marked. When no passage clears the threshold, Folio says there is no direct answer in scope and points to the collection that probably holds it. Behind the answers, Collections shows every file becoming searchable page by page (with OCR for scanned PDFs), Evaluations runs a golden set of questions against each settings change and shows why each failure failed, and Settings lets you move the threshold on a test question before anything is saved.

FIG. 02 / HOW IT WORKS

How it works

Step 1: Ask
Ask in plain words over the collections you pick. Folio pulls the best passages with their relevance scores and marks the words that matter; then the answer is set one sentence at a time, and each numbered citation appears only after its passage. Click a citation and the source page opens with the cited block bracketed. When nothing clears the threshold, Folio says so and suggests where to look.
Step 2: Collections
Every collection lists its files, chunks, owner and who can read it. A new contract becomes searchable page by page: text is extracted, chunked, embedded and indexed, a page map fills as each page is embedded, and the newest chunks stack up beside it. A scanned PDF with no text layer is flagged, and one click queues OCR for it.
Step 3: Evaluations
A golden set of questions runs against the current settings and is compared with the previous run: correct answers, faithfulness, recall, citation precision, abstentions and latency, each with its change. Open a failure and Folio shows why it failed — the expected passage ranked just under the threshold, a superseded document cited, a source still indexing — and what to try next.
Step 4: Settings
Tune how Folio finds and cites evidence: the keyword and vector weighting, how many candidates it fetches, the reranker and the citation threshold. A test question runs as you move the threshold: the line slides through the ranked passages and you see which ones now reach the answer model. Saving can require an evaluation run first, so no change reaches people untested.

FIG. 03 / SCREENS

The product, screen by screen

Folio Ask page: the collections to search and an indexing queue on the left; a question about expensing a rental car abroad answered in three sentences with numbered citations, and a follow-up answered “No direct answer in scope”; on the right, the three retrieved passages with relevance scores and the cited block marked on its source page.

Folio Collections page: a catalogue of six collections, a contracts collection's files with pages, chunks and status — one embedding at 64%, one scanned file failed with no text layer — and a panel for the embedding file with its pipeline stages, a page map filling page by page and the latest chunks.

Folio Evaluations page: run #48 compared with run #47 on a 120-question golden set, six metrics with their changes, verdict bars, correct answers by collection and a table of failed questions; one failure is open with its claims and eight ranked passages, the expected one just under the threshold line.

Folio retrieval settings: the pipeline from query through hybrid search, rerank and threshold to answer; sliders for keyword and vector weighting and for the citation threshold at 0.60, an abstain switch and the change history; beside them a test question with eight ranked passages and the threshold line after the fifth.

Folio Ask on a phone: the question, three source slips with their scores, the answer with numbered citations, and the follow-up's “No direct answer in scope” card with a button to add the suggested collection and ask again.

Folio Collections on a phone: chunk, indexed and needs-OCR counts, the file being embedded with its stage rail and page map, a scanned file with a Run OCR button, and the most recently indexed files.

Folio Evaluations on a phone: six metrics with their changes against the previous run, verdict bars for both runs, and the failed questions with what changed and why.

Folio Settings on a phone: the retrieval pipeline, the citation threshold slider, the abstain switch and a test question's ranked passages with the threshold line.

FIG. 04 / DESIGN DIRECTION

Design direction

Library archive: light, quiet and exact. Spectral, a reading serif, carries what is read — answers, passages, titles, metrics — and Instrument Sans what is operated: navigation, tables, buttons, labels. Hairline rules, near-square corners and a single soft shadow, kept for document page sheets. One ink-blue accent marks citations, scores and actions, a pale blue highlight marks cited words, and rust is kept for failures and regressions.

Palette

Paper#F2F0EA
Ink#1E2422
Ink blue#3B5BDB
Highlight#DCE2F7
Rust#A4441C

Type

Spectral: What is read: answers, passages, titles, metrics
Instrument Sans: What is operated: navigation, tables, buttons, labels

FIG. 05 / MOTION LANGUAGE

Motion language

“Evidence first — pull, mark, cite”: passages slide out of the drawer like index cards, a highlighter marks the words that matter, and only then does a sentence, a citation numeral or a verdict appear. Around it the archive keeps its ledger: rules draw left to right, rows are entered one at a time, pages are embedded leaf by leaf. Whole sentences are set, never streamed letter by letter, and nothing glows or bounces.

Orrery

Live

A knowledge-graph extractor that keeps its receipts: people, organizations, places and ideas pulled from your documents, every relation tied to its sentence.

orrery-ie modelNeo4j exportJSON · CSV export