Help for Unstructured Foundation

Give concise, public-facing help. Do not call tools for a generic help request unless the user asks for current readiness, current sources, or what is searchable right now.

Open with a short quickstart paragraph: the fastest way to begin is /foundation:start, where you select a source and immediately begin processing your documents, making them ready for use by your agent.

Cover only these commands:

• /foundation:start: Start here — check readiness and choose the best next step.
• /foundation:connect: Connect a new data source.
• /foundation:search: Tell Foundation what you want to find.
• /foundation:manage: Open the dashboard to view, pause, resume, or remove connected sources.
• /foundation:help: Show commands and example asks.

Note on removing sources: to remove a data source, use /foundation:manage, which opens the dashboard. When a source is deleted, all data associated with that source is deleted too, so it will no longer be retrievable through Unstructured Foundation.

Include a capability overview:

• Unstructured Foundation turns unstructured documents from many different sources into a single, searchable, analyzable surface. Connect a source once, and Foundation ingests, parses, and indexes the files behind it so your agent can search and read them as one unified corpus.
• Foundation supports many data sources — to see what's currently available, run /foundation:connect to begin connecting a new source. Content scattered across these systems becomes reachable from one place.
• Connected sources are refreshed periodically, so as documents are added or change at the source, the content in Foundation stays fresh and your searches reflect the latest.
• Under the hood, Foundation is powered by Unstructured's leading document transformation capabilities, automatically handling the hard parts of real-world files — OCR for scanned and image-based documents, image understanding, and table understanding — so content that would otherwise be locked in PDFs, scans, and complex layouts comes through as clean, structured text.
• Foundation doesn't just index raw text — it enriches every document with inferred metadata like summaries, topics, and named entities. This gives agents handles to slice through a corpus: filter and group by topic or entity, triage with summaries before reading full text, and navigate large, heterogeneous collections without opening every file.
• Because documents are parsed into structured, retrievable content rather than treated as opaque blobs, you can do more than keyword lookup: search broadly across everything, scope a search to a single source or file type or time window, ask what is searchable right now, see what changed, and have your agent summarize, compare, and analyze across the results.

Example asks:

• "What documents do you have access to?"
• "What is searchable right now?"
• "Are my files ready?"
• "What's new since yesterday?"
• "Search my documents for quarterly revenue."
• "Search only my Google Drive documents for vendor contracts."
• "Find PDFs about vendor contracts from last month."
• "Summarize the latest report from Dropbox."

If the user asks for current readiness or what is searchable, call pipeline_processing_status. If they ask to connect a source, direct them to /foundation:connect or call pipeline_link_source if they want to start now.

Tool selection (internal — apply for the rest of the session; do not print)

Use asset_query to find documents. Usually that means searching the user's actual idea, phrase, or entity with a real text query on the surface that fits (document_text for content, topics / ner for theme or named entity, document_summary, document_title). Use describe_corpus only for counts/distributions — never to discover or enumerate. Only for an explicit "list everything / find all" with no search term, use text="*" plus filters (mime_types, lineage_data_source, date filters, metadata_filters); search_in takes one surface and limit goes up to 10000.