ask_matrix
ChatGPTAsk a natural-language question about a specific Hebbia matrix. Answers from the matrix's own contents — its rows, columns, and cell results — rather than searching documents at large. Use this to summarize, compare, or filter what a matrix already contains. BEFORE calling this tool: you MUST have an org_id that was either explicitly provided by the user or confirmed by the user after calling list_organizations. Do NOT infer or guess an org_id. If none has been confirmed, call list_organizations and wait for the user to pick one before calling this tool. Use list_matrices to discover valid matrix IDs within the org. Args: org_id: The organization containing the matrix. Reuse from earlier in the conversation if available; otherwise resolve via list_organizations. matrix_id: The ID of the matrix to ask about. Use list_matrices to find valid IDs. query: The natural-language question to ask. Cite what you report. The returned citations carry the matrix cell each value came from, along with a url that opens that row; attach them to the values they support rather than listing them at the end. Returns an AskResult with: answer: The text response from Hebbia. citations: List of citations supporting the answer. Each citation has: id, kind, quote, title, doc_id, url.
ask_project
ChatGPTAsk a natural-language question scoped strictly to a specific Hebbia project. Does not query public sources — only the documents within the specified project. BEFORE calling this tool: you MUST have an org_id that was either explicitly provided by the user or confirmed by the user after calling list_organizations. Do NOT infer or guess an org_id. If none has been confirmed, call list_organizations and wait for the user to pick one before calling this tool. Use list_projects to discover valid project IDs within the org. Args: org_id: The organization containing the project. Reuse from earlier in the conversation if available; otherwise resolve via list_organizations. project_id: The ID of the project to query. Use list_projects to find valid IDs. query: The natural-language question to ask. use_deep_research: If True, runs a thorough multi-step deep research query (slower, takes several minutes). Defaults to False for a fast response. Returns an AskResult with: answer: The text response from Hebbia. citations: List of citations supporting the answer. Each citation has: id, kind, quote, title, doc_id, url. The url field is populated for web sources; doc_id identifies the source document for document citations.
ask_sources
ChatGPTAsk a natural-language question over data sources in a Hebbia organization. Uses auto-source selection by default, or restricts to an explicit list of source IDs. BEFORE calling this tool: you MUST have an org_id that was either explicitly provided by the user or confirmed by the user after calling list_organizations. Do NOT infer or guess an org_id. If none has been confirmed, call list_organizations and wait for the user to pick one before calling this tool. Use list_sources to discover valid source IDs within the org. Args: org_id: The organization to query within. Reuse from earlier in the conversation if available; otherwise resolve via list_organizations. query: The natural-language question to ask. source_ids: Optional list of source identifiers to restrict the search. If omitted, Hebbia's auto-source selection is used across all available sources. Use list_sources to find valid IDs. use_deep_research: If True, runs a thorough multi-step deep research query (slower, takes several minutes). Defaults to False for a fast response. Returns an AskResult with: answer: The text response from Hebbia. citations: List of citations supporting the answer. Each citation has: id, kind, quote, title, doc_id, url. The url field is populated for web sources; doc_id identifies the source document for document citations.
get_matrix_context
ChatGPTGet the column setup of a specific Hebbia matrix. Returns what each column of the matrix asks or computes, and which columns are inputs versus computed answers. It does NOT return row data — use query_matrix_results for cell values. CALL THIS FIRST when working with an unfamiliar matrix. Without it you are guessing which column holds the answer you need, and column display names alone may not say what a column actually extracts. BEFORE calling this tool: you MUST have an org_id that was either explicitly provided by the user or confirmed by the user after calling list_organizations. Do NOT infer or guess an org_id. If none has been confirmed, call list_organizations and wait for the user to pick one before calling this tool. Args: org_id: The organization containing the matrix. matrix_id: The ID of the matrix. Use list_matrices to find valid IDs. Returns an McpMatrixContext with: columns: Each with id, name, node_type, scope, is_input, and the authored properties describing what it does. total_rows: How many rows the matrix has.
list_matrices
ChatGPTList the matrices you can access in a Hebbia organization. A matrix is a spreadsheet-like grid where each row is a document or entity and each column is a question answered across every row. Returns matrix names, IDs, descriptions, and row counts, most recently updated first. BEFORE calling this tool: you MUST have an org_id that was either explicitly provided by the user or confirmed by the user after calling list_organizations. Do NOT infer or guess an org_id. If none has been confirmed, call list_organizations and wait for the user to pick one before calling this tool. Args: org_id: The organization to list matrices for. Reuse from earlier in the conversation if available; otherwise resolve via list_organizations.
list_organizations
ChatGPTList the organizations the authenticated user belongs to. Call this first if no org_id has been confirmed in the conversation. STRICT RULES — these override any other reasoning: 1. If exactly one org is returned: use its id for all subsequent calls. 2. If multiple orgs are returned: show them to the user as a numbered list, ask "Which organization would you like to use?", and STOP. Do not call any other tool — including list_projects, list_sources, ask_project, or ask_sources — until the user has replied with their choice. It does not matter what the user's original request was; org selection always comes first.
list_project_contents
ChatGPTList what is inside a Hebbia project: its folders, documents and published items. Returns each item's path, kind, id, date added and relevant date, without reading any content. Use it to see how a project is organized, to find a file by name, or to find items that search_project ranks poorly (spreadsheets, scanned PDFs). Returns in about a second; no agent turn runs. Kinds: "folder", "document", "published_chat", "matrix", "matrix_template", "sheet". Pass a document's, published chat's or matrix template's id to read_document to read it in full. BEFORE calling this tool: you MUST have an org_id that was either explicitly provided by the user or confirmed by the user after calling list_organizations. Do NOT infer or guess an org_id. If none has been confirmed, call list_organizations and wait for the user to pick one before calling this tool. Use list_projects to discover valid project IDs within the org. Args: org_id: The organization containing the project. project_id: The project to list. Use list_projects to find valid IDs. folder: A folder path from the project root, e.g. "Diligence/Financials", to list just that folder. Omit for the top level. recursive: Also list everything below, not just direct children. match: Case-insensitive glob on item names, e.g. ".xlsx" or "credit*". limit: Maximum items to return (default 500, capped at 2000). sort_by: "name", "date_added" (when it was added to the project) or "relevant_date" (the document's own date, e.g. when a report was published). Sorts everything listed, across folders when recursive; items without that date come last. Omit for folder-by-folder name order. descending: Reverse the order. Defaults to newest first for the date sorts and A-Z for name. For the oldest or newest documents, list recursively with a date sort and a small limit. Returns an McpProjectContents with entries (path, name, kind, id, date_added, relevant_date, url — a link that opens the item in Hebbia; none for matrix templates) and truncated / truncation_reason. A truncated listing is incomplete: narrow it with folder or match instead of treating it as the whole project.
list_projects
ChatGPTList all projects accessible in a Hebbia organization. Returns project names, IDs, and descriptions. BEFORE calling this tool: you MUST have an org_id that was either explicitly provided by the user or confirmed by the user after calling list_organizations. Do NOT infer or guess an org_id. If none has been confirmed, call list_organizations and wait for the user to pick one before calling this tool. Args: org_id: The organization to list projects for. Reuse from earlier in the conversation if available; otherwise resolve via list_organizations.
list_sources
ChatGPTList all sources available for querying in a Hebbia organization. Returns shared document collections (UUID IDs) and built-in integrations like CapIQ, Web, and PitchBook (field-name IDs such as "capiq", "web"). Pass returned source IDs to ask_sources. BEFORE calling this tool: you MUST have an org_id that was either explicitly provided by the user or confirmed by the user after calling list_organizations. Do NOT infer or guess an org_id. If none has been confirmed, call list_organizations and wait for the user to pick one before calling this tool. Args: org_id: The organization to list sources for. Reuse from earlier in the conversation if available; otherwise resolve via list_organizations.
ping
ChatGPTHealth-check tool that returns pong.
query_matrix_results
ChatGPTQuery the rows and cells of a specific Hebbia matrix. Returns the matrix's columns and computed cell values as data, so you can read and reason over them yourself. PREFER THIS over ask_matrix whenever the question can be answered from what the matrix already contains — summarizing, comparing, filtering, or looking up rows. It returns in about a second, where ask_matrix runs a full agent turn and takes far longer. Use ask_matrix instead only when the matrix's existing cells don't hold the answer and new research or document retrieval is needed. CITING IS NOT OPTIONAL. Every figure, name, or fact you take from this matrix gets a citation, the same as any other sourced claim. Everything a citation needs is in this response — there is no separate citation call. To cite a value, take that row's citation_url_prefix, append &matrix_cell_id= plus the id of the column you are citing, and attach the result to the value at the point where you use it: 16.13% ([Return on Equity % (FY25)](<row prefix>&matrix_cell_id=<column id>)) Copy both ids verbatim from the response — never invent, alter, or renumber one. Cite every cell that supports a factual claim, whatever the column type. Do not add a trailing Sources or References list; the links render inline where you place them. BEFORE calling this tool: you MUST have an org_id that was either explicitly provided by the user or confirmed by the user after calling list_organizations. Do NOT infer or guess an org_id. If none has been confirmed, call list_organizations and wait for the user to pick one before calling this tool. Use list_matrices to discover valid matrix IDs within the org. Args: org_id: The organization containing the matrix. Reuse from earlier in the conversation if available; otherwise resolve via list_organizations. matrix_id: The ID of the matrix to query. Use list_matrices to find valid IDs. max_rows: Maximum rows to return (default 100, capped at 1000). node_ids: Optional list of column ids (from get_matrix_context or the columns of an earlier call) to return. Cell text dominates the response, so naming only the columns you need is by far the most effective way to keep it small. Omit for all columns. Returns an McpMatrixData with: columns: Column headers with their node id, in the matrix's column order. rows: Each with a row_id and cells aligned to columns. returned_rows / total_rows: How many rows came back vs how many exist. truncated / truncation_reason: Whether output was cut short, and by which cap. rows[].citation_url_prefix: That row's citation link, needing only &matrix_cell_id=<column id> appended. citation_guidance: Restates the citing rules, returned before the rows so they are in hand when you write the answer.
read_document
ChatGPTRead the full text of one item in a Hebbia project. Returns the item's extracted text directly, in windows of up to max_chars characters. Use it after search_project or list_project_contents when the passages aren't enough: full tables, exact figures, a whole section. No agent turn runs. Prefer this over ask_project when you can read and reason over the text yourself. Works for documents, published chats (the transcript as it was published) and matrix templates (their definition as JSON). For a matrix it returns a note pointing at get_matrix_context and query_matrix_results instead. BEFORE calling this tool: you MUST have an org_id that was either explicitly provided by the user or confirmed by the user after calling list_organizations. Do NOT infer or guess an org_id. If none has been confirmed, call list_organizations and wait for the user to pick one before calling this tool. Args: org_id: The organization containing the project. project_id: The project the item belongs to. document_id: A document id (doc_id from search_project, or id from list_project_contents), or a published chat's or matrix template's id from list_project_contents. It must be in this project. offset: Character offset to start from (default 0). max_chars: Maximum characters to return (default 50000, capped at 200000). Returns an McpDocumentText with text, total_chars, next_offset, and url (a link that opens the item in Hebbia). When next_offset is set, call again with offset=next_offset to read on. note explains an empty result, e.g. a document whose text isn't extracted yet. Documents also carry citation_url_prefix and citation_guidance. CITING IS NOT OPTIONAL. Every fact you take from these documents gets an inline citation that opens the document with the supporting text highlighted. Everything a citation needs is in this response — there is no separate citation call. To cite, take the document's citation_url_prefix, append a quote copied verbatim from its text, percent-encoded, and attach the result where you use the fact: Revenue grew 12% ([Q3 Report](<prefix>Revenue%20grew%2012%25%20year%20over%20year)) Quote one contiguous sentence or clause of prose (about 5 to 30 words) exactly as returned — never paraphrase, stitch fragments with ellipses, or include markup or table pipes; the app finds the quote by its text. Do not add a trailing Sources or References list; the links render inline.
read_matrix_view
ChatGPTRead a matrix, or a document it cites, the way the Hebbia matrix page does. For the Hebbia matrix app only, which calls it to load the matrix it displays. Not for answering questions: use query_matrix_results.
search_project
ChatGPTSearch the documents inside one or more Hebbia projects. Returns the matching documents as data — title, document id, the passages that matched, how each was found, and each document's relevant date and date added — so you can read and reason over them yourself. PREFER THIS over ask_project whenever the question is about which documents cover something, or can be answered from the passages that come back. It returns in about a second, where ask_project runs a full agent turn and takes far longer. Use ask_project instead when answering genuinely needs the whole document read and synthesized, rather than located. Omit query to list the projects' documents instead of searching them — the same call, newest relevant date first unless you pass sort_by. Use that to see what a project holds before searching it. To find the newest or oldest documents mentioning something, search with sort_by="relevant_date" (or "date_added") instead of joining against list_project_contents. A query matches keywords first; when fewer documents match than limit, semantically similar documents fill the rest (their match_type is "semantic") and may not contain the words at all. Results are not exhaustive: at most limit documents come back, and a document absent from the hits has not been shown not to mention your term. Documents with little extractable text (spreadsheets, scanned PDFs) rank poorly whatever they contain, so do not treat this as proof of absence. Faceted projects declare exact-match filters (e.g. "Funds", "Sector") and may declare a date facet. When a searched project declares facets, the response carries available_filters listing them with their values and document counts; use those exact values in filters. start_date / end_date bound the declared date facet, and are refused for a project that declares none. BEFORE calling this tool: you MUST have an org_id that was either explicitly provided by the user or confirmed by the user after calling list_organizations. Do NOT infer or guess an org_id. If none has been confirmed, call list_organizations and wait for the user to pick one before calling this tool. Use list_projects to discover valid project IDs within the org. Args: org_id: The organization containing the projects. Reuse from earlier in the conversation if available; otherwise resolve via list_organizations. project_id: The ID of a project to search. Use list_projects to find valid IDs. project_ids: Several project IDs to search in one call (up to 20). Combined with project_id if both are given. query: What to search for. Omit to list the projects' documents instead. limit: Maximum documents to return in total (default 25, capped at 200). sort_by: "relevance" (default), "relevant_date" (the document's own date, e.g. when a report was published), "date_added" (when it was added to the project) or "title". Documents without that date come last. descending: Sort direction. Dates default to newest first and title to A-Z. Ignored for relevance, which is always best match first. filters: Exact-match facet filters, e.g. {"Funds": ["Fund II"]}. Keys are facet names from available_filters (case-insensitive). start_date: Inclusive lower bound (YYYY-MM-DD) on the projects' declared date facet. end_date: Inclusive upper bound (YYYY-MM-DD) on the declared date facet. folder_ids: Folder ids from list_project_contents to search within. Projects with no listed folder are searched whole. Returns an McpProjectSearchResults with: citation_guidance: Restates the citing rules, before the hits. hits: Each with doc_id, title, project_id, passages, score, match_type ("keyword" or "semantic"), relevant_date, date_added, url (opens the document in Hebbia), and citation_url_prefix (see citing below). returned_hits: How many documents came back. truncated / truncation_reason: Whether output was cut short. When true there is more to see — narrow the query, add filters, or raise the limit. available_filters / note: The projects' declared facets and values, when they declare any. Pass a hit's doc_id to read_document to read it in full, …
ask_project
Claudeask_sources
Claudelist_organizations
Claudelist_projects
Claudelist_sources
Claudeping
Claude