Common questions

The free tier includes 1 project, 500 MB of storage, 2,000 queries per month, 25 research credits per month, and 500 ingestion units per month. Research at quick and balanced depth is available. No credit card required.

Free accounts are suspended after 90 days of inactivity. Your data is retained; the account is simply paused until you return.

Research credits are consumed by /research requests based on depth: quick costs 0.5 credits, balanced costs 1 credit, deep costs 3, and exhaustive costs 8. Add a BYOK key to bypass credits entirely — you pay your model provider directly with no Dewey markup.

Credits reset at the start of each billing cycle and do not roll over.

BYOK (bring your own key) lets you connect your own OpenAI or Anthropic API key to Dewey. When a BYOK key is configured, Dewey uses it for all research requests at every depth. No research credits are consumed, and Dewey adds no markup on generation — you pay your provider at cost.

At exhaustive depth with Claude Opus 4.6, Dewey achieved 87.3% accuracy on FinanceBench. The typical per-question cost in API tokens for that run was a few cents. See the study.

On Pro and Team, overages are on by default: additional queries cost $1.00 per 1,000, and additional ingestion units cost $0.008 each. You can disable overages for your org from the Usage section of the dashboard — once disabled, requests above your monthly limit return a 429 with an upgrade prompt instead of incurring charges.

Yes. You can upgrade, downgrade, or cancel at any time from the Usage section of your dashboard. Upgrades take effect immediately. Downgrades and cancellations take effect at the end of the current billing period.

The free tier is the trial. It gives you real access to the API, real retrieval, and 25 research credits per month at any depth. Upgrade to Pro when you need higher limits, BYOK, or full model flexibility.

Yes. Pro and Team plans can be purchased monthly or annually. Annual billing saves 20% — $39/month for Pro ($468/yr) and $159/month for Team ($1,908/yr). Use the Monthly / Annual toggle on the pricing page or the upgrade panel in your dashboard to choose your billing interval.

PDF, DOCX, PPTX, JSON, HTML, Markdown, and plain text. More formats are added regularly. If you have a specific format you need, let us know.

A typical 10-page PDF is fully processed and ready to query in under a minute. Large documents, image-heavy PDFs, or PDFs with complex tables take longer, particularly if AI captioning is enabled.

You can subscribe to real-time status events via SSE to know exactly when a document is ready, without polling.

Ingestion units measure the processing work required to ingest a document. PDFs and PowerPoint files consume one unit per page (minimum one). All other formats — DOCX, HTML, Markdown, and plain text — consume one unit regardless of length. Units reset at the start of each billing cycle.

At Pro and above, Dewey can run a vision model over images, figures, and diagrams embedded in your documents to generate text descriptions, and an LLM over tables to generate structured captions. Both are indexed as searchable chunks alongside the document text.

This makes charts, graphs, and tabular data fully retrievable by semantic search, not just the text that surrounds them. Captioning requires a BYOK key.

Yes. Send a DELETE /v1/documents/:id request and the document, its chunks, embeddings, and section summaries are permanently removed. Deletion is immediate and irreversible.

/query is a single-pass hybrid search: it retrieves the most relevant chunks for a query and returns them ranked, with document and section provenance. It is fast and predictable.

/research runs a multi-step agentic loop. The model issues multiple search calls, reads results, decides whether to dig further, and synthesizes a final cited answer in Markdown. It is slower but can answer questions that require finding and combining information from multiple locations in a corpus.

Depth controls how many tool calls the research agent can make before producing its answer:

• quick : 1 to 3 searches. Best for targeted factual lookups.
• balanced : up to 10 searches. The default for most Q&A and agent workflows.
• deep : up to 25 searches. For synthesis across large or multi-document corpora. Costs 3 credits per run.
• exhaustive : up to 50 searches. For multi-hop reasoning, complex analysis, and full corpus traversal. Costs 8 credits per run.

Yes. Every /research response includes a sources array listing the document, section, and chunk each part of the answer was drawn from. Citations are not optional and cannot be disabled.

At Pro and above, you can use any model from OpenAI or Anthropic via your own API key. Pass the model identifier in the model field of the request. Dewey manages the tool loop and context assembly; the model provides the reasoning.

Documents, embeddings, and metadata are stored on US-based infrastructure. Enterprise plans include a private deployment option if you need data residency in a specific region or a dedicated environment.

Yes. Data is encrypted at rest and in transit. All API traffic uses TLS 1.2 or higher.

No. Your documents are used only to serve your queries. They are never used to train Dewey models or shared with any third party.

Individual documents can be deleted via the DELETE /v1/documents/:id endpoint. Entire collections can be deleted via DELETE /v1/collections/:id, which removes all documents, embeddings, and metadata in the collection. For bulk export, contact hello@meetdewey.com.

A DPA is available for Enterprise customers. Contact enterprise@meetdewey.com to request one.

Claim extraction runs an LLM over each document and pulls out discrete, atomic facts — things the document actually asserts. Each claim is scored by importance on a 1–5 scale (1 = minor detail, 5 = critical finding) and linked back to the exact section it came from. Claims are the foundation for contradiction detection and can also be browsed, filtered, and searched directly in the dashboard.

Once claims are extracted, contradiction detection compares them across the entire collection, clusters statements that conflict with each other, assigns a severity (high / medium / low), and generates a plain-English explanation plus a suggested resolution instruction. That instruction can be applied directly to the collection so future research answers respect the resolution.

Detection runs asynchronously — for large collections it may take a few minutes. You can monitor progress in the Runs tab on the collection page.

Open the Contradictions tab on a collection, expand a contradiction, and click Apply resolution. Dewey appends the suggested instruction to the collection's research instructions so it takes effect immediately for all future research queries. You can revert it at any time by clicking Revert.

Neither feature consumes processing units. Both require a BYOK (bring-your-own-key) API key — you pay your model provider directly at cost, and Dewey charges nothing for the runs. Claim extraction is available on Pro and Team plans.

High — the conflicting claims make materially different assertions that could produce incorrect or misleading research answers. Medium — claims are in tension but the discrepancy may be contextual or temporal. Low — minor inconsistency unlikely to affect research quality. All severity levels are surfaced in the UI; filter by severity to focus on what matters most.

Dewey measures how much content two documents share by looking for overlap across their existing chunks, then clusters documents whose overlap is high enough in both directions. This catches true duplicates as well as near-duplicates — e.g. a preprint and its final journal version, a PDF and its OCR'd reupload, or the same report mirrored across two archives. No LLM calls are involved, so detection is fast and consumes no processing units.

Each cluster has one canonical document — by default, the earliest-uploaded member. The rest are marked as near_duplicate and excluded from retrieval, research, and contradiction detection. This prevents research answers from citing the same content under three different filenames and stops contradiction detection from flagging a document as conflicting with itself. The near-duplicates are not deleted — they remain visible in the dashboard and can be promoted to canonical or disbanded at any time.

Deduplication is off by default. Enable it on a per-collection basis via the Enable deduplication toggle in collection settings, or through PATCH /collections/:id with enableDeduplication: true. Once enabled, trigger a detection run from the Duplicates tab or via dewey_detect_duplicates in the MCP server.

Yes. Open the Duplicates tab, expand a group, and click Promote next to any member. The old canonical becomes a near-duplicate and the new canonical re-enters retrieval immediately. You can also disband a group entirely if the clustering was wrong — all members rejoin retrieval as distinct documents.

Nothing. Detection reuses the chunk embeddings Dewey already computed at ingest time, makes no LLM calls, and consumes no processing units. Run it as often as you like.

MCP (Model Context Protocol) is an open standard for exposing tools to AI assistants. The Dewey MCP server makes your document collections available as tools inside Claude, Cursor, and any other MCP-compatible agent with no custom integration required.

Add https://mcp.meetdewey.com/mcp as a custom MCP server in your client and authorize in the browser. Your documents become searchable from within the assistant immediately — no API key to copy.

Yes. The dewey CLI is a single Go binary that wraps the same REST API as the SDKs. Install it on macOS or Linux with:

curl -fsSL https://raw.githubusercontent.com/meetdewey/dewey-cli/main/install.sh | sh

The script detects your OS and architecture, verifies the SHA-256 checksum, and installs to ~/.local/bin (no sudo required). Set DEWEY_API_KEY and you can upload documents, run hybrid search, stream cited research, and invoke hosted agents from any shell. Every command supports --json for piping into other tools.

For Windows, download dewey_<version>_windows_amd64.zip from the Releases page and extract dewey.exe to a directory in your PATH.

Yes. Official client libraries are available for TypeScript and Python. Both wrap the REST API and handle authentication, retries, and SSE streaming.

https://api.meetdewey.com/v1. Full reference documentation is at meetdewey.com/docs.

Rate limits are based on your plan's monthly query and document quotas. There are no per-second or per-minute burst limits under normal usage. If you need higher throughput for a batch workload, contact us.