Productivity

Deep Document Search

by Forest AI

Overview

Deep Document Search turns ChatGPT into a powerful interface for your private document knowledge base. Search and analyze PDFs, Word files, scanned documents, and large document collections using natural language; find relevant information across thousands or millions of files; and get accurate, traceable answers with citations back to the original documents. It is built for document-intensive workflows such as legal research, regulatory compliance, corporate knowledge, due diligence, research, and document analysis.

Tools

batch_get_item

ChatGPT

Retrieve the full text content of one or more items by their item_ids (returned by search_documents). An item groups one or more documents together â€” a document is a type of item. Use this when the user asks to read the complete content of specific documents/items found via search. Content is fetched from the item content store in a single batch; pass up to 20 item_ids.

describe_dataset

ChatGPT

Inspect a dataset: stable field IDs, types, dimension/measure roles, row count, and supported chart types. Use field IDs (never display names) in run_analysis.

get_chart

ChatGPT

Load a saved chart (exploration) definition and revision with a first-party view link.

get_query_result

ChatGPT

Poll a query job: status plus bounded typed rows, columns, truncation flag, dataset version, and provenance when terminal.

list_datasets

ChatGPT

List analytics-ready datasets in the organization. Returns dataset IDs and names with conversion status. Resolve names to IDs here before calling other tools.

list_projects

ChatGPT

List projects in your organization. Optionally pass workspace_id to see only the projects inside that workspace, and/or project_name to fuzzy-match only the projects whose name contains that text (case-insensitive substring match, e.g. "annual" also finds "Annual Budget"). Returns project IDs, names, and workspace IDs. Use the IDs to scope a document search (pass them to search_documents via project_ids).

list_workspaces

ChatGPT

List workspaces in your organization. Optionally pass workspace_name to fuzzy-match only the workspaces whose name contains that text (case-insensitive substring match, e.g. "fin" also finds "Finance"). Returns workspace IDs and names. Use the IDs to scope a document search (pass them to search_documents via workspace_ids) or to filter projects (pass one to list_projects as workspace_id).

preview_chart

ChatGPT

Validate a visualization spec against a completed job's columns and return the portable chart spec. Requires chart IDs from describe/get results, never invented names.

profile_dataset

ChatGPT

Bounded column statistics (min/max/avg/distinct counts/null rates). Never returns raw rows. Use it to pick sensible filters and aggregations.

save_chart

ChatGPT

Persist a chart definition (query + visual spec) under a name. Returns the chart ID, revision, and first-party view URL.

search_documents

ChatGPT

Keyword search across the user's uploaded documents. Returns relevance-ranked results with scores, highlighted matching passages (snippets), item IDs, and match fields (description, mime type, workspace/project). Provide one or more short, specific phrases via the phrases array (e.g. 2-4 key words, not full sentences) â€” documents matching any phrase are returned; exact phrase matches are tried first, and if nothing matches the tool retries by matching individual words. You can control how much context you get by setting highlight_length (characters per highlighted snippet, default 450) â€” use a larger value (e.g. 2000-4000) when you need more surrounding information to answer a detailed question, and num_snippets to control how many snippets per result are returned (default 3). You can restrict results by passing arrays of item_ids, project_ids, or workspace_ids: within each array IDs are combined with OR (match any), and the groups are combined with AND (e.g. (item1 OR item2) AND (project1 OR project2)).