Turn your docs into
answers with citations
Upload any collection of documents: reports, research, contracts, policies. Ask questions in plain English and get cited answers drawn from the exact passages. Build it into your app via REST API, or connect it to Claude and Cursor in minutes.
No credit card required · Free tier available · Use with Claude, ChatGPT, or API
In the best-fit poliovirus transmission model for the 2010 Tajikistan outbreak, the reproduction number for older children and adults at the start of the outbreak was estimated as 0.46 (95% CI: 0.42–0.52).
Critical“Reproduction number for older children and adults at the start of the outbreak* | 0.46 (0.42 - 0.52)”
Blake2014-PNAS.pdf › Discussion
In the Republic of Congo outbreak, an estimated 75% of all susceptible individuals were adults.
Critical“This was not simply the result of the large number of susceptible adults (an estimated 75% of all susceptible individuals)”
Blake2014-PNAS.pdf › SI Appendix
In developing countries, only 73% of children have detectable antibody to poliovirus type 1 after three doses of trivalent OPV, with a range of 36%–99%.
Important“only 73% (range, 36%–99%) ... of children in developing countries have detectable antibody to poliovirus types 1”
Patriarca1991 - factors affecting immunogenicity of OPV.pdf › Immunogenicity in Developing Countries
The two claims describe the 24-week OPV challenge in what appears to be the same combined IPV/OPV study, but they identify different monovalent challenge serotypes: one says monovalent OPV type 3, while the other says monovalent type 1 OPV.
The 24-week challenge dose was monovalent OPV type 3.
“At the 24-week visit, the children received a dose of monovalent OPV type 3 as a challenge.”
ParentChatelet2003 - combined OPV IPV.pdf›2.2. Study design
The 24-week challenge dose was monovalent type 1 OPV.
“Infants were challenged at 24 weeks of age with a dose of 10⁶ TCID₅₀ of monovalent type 1 OPV...”
WHO1997 - combined IPV OPV in Oman Gambia Thailand.pdf›Methods
When summarizing the 24-week challenge dose in the combined IPV/OPV infant studies, do not state a single serotype without checking the original trial reports. Parent du Châtelet 2003 reports type 3; the 1997 WHO report describes type 1. Verify whether these are different studies or a citation inconsistency before answering.
CANONICAL
Kew2002 - outbreak of poliomyelitis in hispaniola associated with circulating type 1.pdf
Kew et al._2002_Outbreak of poliomyelitis in Hispaniola associated with circulating type 1 vaccine…
Coverage to canonical: 100% · From canonical: 89%
Kew2002-Outbreak of Polio in Hispaniola Associated with CVDPV1.pdf
Coverage to canonical: 93% · From canonical: 93%
At exhaustive depth, Dewey surpasses the full-document-in-context baseline on FinanceBench, a 150-question benchmark of financial Q&A over real SEC filings. See the study →
Features
Every piece of the pipeline.
None of the plumbing.
Upload any document
PDFs, Word files, PowerPoint decks, web pages, plain text. Dewey converts them to clean, searchable content automatically. No preprocessing or pipelines to maintain.
Ask questions, get cited answers
Ask a plain-language question and get an answer drawn directly from your documents. Every response links back to the exact section it came from. No hallucinations, no guessing.
Choose your depth
A quick lookup or a thorough multi-document research session, all from the same endpoint. Four depth levels let you balance speed against thoroughness. At exhaustive depth, Dewey reaches 83.7% accuracy on FinanceBench, compared to 19% for standard vector RAG. Deep and exhaustive require Pro and a BYOK key.
See the study →Works with Claude and Cursor
The Dewey MCP server puts your document collections directly inside Claude, Cursor, and any MCP-compatible AI assistant. Your knowledge base, wherever you already work.
Structure preserved, not flattened
Dewey parses the actual heading hierarchy of every document — sections are first-class API objects, not anonymous text fragments. Query at the chunk, section, or document level and get citations by section name with exact character offsets.
Your models, your bill
On Pro and above, bring your own OpenAI, Anthropic, or Google Gemini key and Dewey uses it across all research depths — including AI image captioning, which makes figures and diagrams searchable alongside your text. No markup on generation, ever.
Hosted agents
Save the agent. Skip the scaffolding.
Most teams end up rebuilding the same agent loop in every script: prompt, tools, retrieval, citation handling. Dewey hosts it for you, so the agent is a first-class object you create once and call from anywhere.
Saved configurations
Package a system prompt, tool set, collection scope, and model into a named agent. Edit once, run forever — no copy-pasting prompts between scripts.
Run from anywhere
Invoke from the dashboard, the REST API, the CLI (`dewey agents invoke`), or the Python and TypeScript SDKs. Same agent, same behavior, anywhere your code runs.
Full run transparency
Every run records the full tool-call trace with inline citations, clickable source rows, and per-tool result rendering. No black-box loops.
sec-filings-analyst
Claude Haiku 4.5 · 2 collections · 3 tools
System prompt
Answer questions about SEC filings. Always cite the section. Prefer scan_sections before retrieval.
Tools
search_collectionscan_sectionsfetch_documentLast run · 1.4s
- →scan_sections("R&D capitalization")
- →search_collection · 4 chunks
- Answer · 3 citations
dewey agents invoke sec-filings-analystRun →How it works
From documents to answers in three steps.
Add your documents
Create a collection and upload your files: PDFs, Word docs, PowerPoint decks, web pages, plain text. Dewey accepts what you have. No conversion or preprocessing on your end.
Dewey does the rest
Every document is automatically converted, structured, and indexed for both keyword and semantic search. Live status events keep your app in sync as each file moves through the pipeline.
Search, ask, and integrate
Search across your corpus, ask questions and get cited answers, or plug your collections into Claude, ChatGPT, and Cursor via the MCP server. All from the same API.
CLI
Dewey from your terminal.
Upload documents, run hybrid search, and stream cited research from any shell. The same API you build against, with a shape your scripts already understand.
$ curl -fsSL https://raw.githubusercontent.com/meetdewey/dewey-cli/main/install.sh | sh# upload a folder of PDFs and watch them index live$ dewey upload ./papers/*.pdf -c research --watch# stream a cited research answer to your terminal$ dewey research research "what are the key findings?" --depth deepOne-line install
Single Go binary. No Python or Node runtime required. Installs to ~/.local/bin by default — no sudo, no PATH gymnastics.
Pipe-friendly by design
Every command supports --json with a stable contract across minor versions. Data on stdout, diagnostics on stderr — drop it into any shell pipeline.
Invoke your hosted agents
dewey agents invoke <name> runs your saved agents from cron jobs, scripts, or CI. Same agent you built in the dashboard, now in your terminal.