AI in Return

Return’s AI is built around one principle: your documents are yours. Local processing is the default, cloud AI is opt-in, and every feature works whether you’re online or off.

There are four pieces that work together. This page is a map - each section below points to its full article.

Providers - where your AI runs

Return supports two ways of running AI: local (through the built-in engine on your own machine) and cloud (Anthropic’s Claude through Return’s proxy). Local is the default; cloud is opt-in through a paid plan. You can switch per conversation.

The trade-off is privacy vs. capability. Local models are smaller but never touch the network; cloud models are more capable but require trust in a third party. Most everyday tasks, summarization, editing, Q&A, work great locally.

→ AI Providers - choose between local and cloud → Local AI (Built-in Engine) - pick a model, switch models, troubleshoot

Search - finding what you need

Return has two search modes: text search finds exact matches instantly, and semantic search finds content by meaning. Search for “payment deadlines” and semantic search will surface a paragraph about “invoices are due within 30 days” - even though the words don’t overlap.

Semantic search runs on embeddings generated locally; nothing leaves your machine.

→ Search - how text and semantic search work

Returns - reusable AI prompts

A Return is a YAML template that captures a repeatable AI task: “summarize this contract”, “extract action items”, “translate to Polish”. You author a Return once, then invoke it from the Quick Actions bar on any document.

Returns can produce text, lists, structured JSON, Traffic Light assessments (green/yellow/red), or inline annotations - depending on the output type you choose.

Large documents

A document too large to analyze in one pass is never rejected. Return splits it into sections along its structure, runs the Return on each section, and merges the results into one answer - so nothing is silently dropped. On a paid plan you see the cost up front: a document that splits into N sections costs about N actions (text and JSON Returns add one more action to merge the sections into a single coherent result). On the local engine it’s always free and unlimited; you just see “processed in N parts”.

Analytical Returns can opt into shared context: before the sections run, a cheap first pass extracts the document’s defined terms and hands them to every section. That lets a finding in a late section reference a term defined on the first page - important for “nothing missed” on long contracts. It costs one extra action and only runs when the document is large enough to split. The built-in Risk Assessment uses it; you can enable it on your own Returns with batch.shared_context: definitions in the YAML.

Chains - composing Returns

When you find yourself running two or three Returns in sequence on the same document, combine them into a Chain. The output of step N becomes the input of step N+1, the user fills in one form covering all variables, and the whole sequence runs as a single Quick Action.

Large documents work inside a chain too: each step runs the same section-and-merge process as a standalone Return, so a big document being analyzed in step 1 is split into sections under that step, and its merged output flows on as a single input to step 2. Because later steps work on already-condensed output, the total cost isn’t fixed - the pre-run estimate shows a range (for example “~5-7 actions”) rather than a single number, and you accept the range before it runs. Local is unlimited, as always.

→ Composing Returns - chain Returns into multi-step workflows

What’s next

If you’re new, start with AI Providers to pick local or cloud. If you want to set up local AI now, jump to Local AI (Built-in Engine). For workflow automation, see Composing Returns.