CiteLLM

Displays answers and claim assessments about PDFs with highlighted source evidence in one interactive viewer, and saves a review identified by review_id. Takes the draft_id of a validation, or the complete claims. One request supports up to 100 claims per PDF and up to 20 PDFs (additional_documents) as document tabs. Each claim takes source quotations, optional page hints (page 1 is the first page of the file, including the cover) and, for a derived value, a calculation with the formula, substituted source values, units and result. Rejects the request without saving a review if a quotation does not match its PDF. insufficient_evidence marks a claim the source does not establish, and follow_up holds an optional question for whoever prepared the document, listed in the evidence report and the Excel export. Matching confirms where text appears, not whether a claim is true. comparison.fields lists the items extracted from each PDF, each with an id and label, for two or more documents. Each claim then sets field, with exactly one claim per field per document; its text is that document's value, and a value the PDF does not state is insufficient_evidence with no quotations. The viewer shows a fields-by-documents table whose cells open the highlighted source passage. requirements lists the requirements that the documents are checked against in a gap analysis, with a title and an id, short label and optional full text for each. Each claim then sets requirement, with exactly one claim per requirement across all documents: supported means covered, partially_supported means partly covered, and insufficient_evidence with no quotations means a gap. A partly covered or gap requirement also sets missing, one sentence on what the documents leave out. The viewer shows a requirements table with status counts and exports a gap report. tie_out ties a draft PDF to its sources, such as a covenant compliance certificate to the credit agreement and financial statements it reports on: session_id is the draft and tie_out.source_session_ids lists the source PDFs, in place of additional_documents. Each claim is then one figure or factual statement in the draft, in draft order, with draft_quote holding the exact draft text and source_session_id naming the source PDF its evidence quotes; a quotation from another source PDF names it in its own source_session_id, so one item can cite several sources, up to 12 quotations per item. A source_session_id may give only the first 8 characters of a session ID that the request gives in full. A session a quotation names that tie_out.source_session_ids leaves out is added as a source. A quotation may also come from the draft itself, under the draft's session_id, away from the text the item ticks: on another page, such as the draft's own schedule, or on the same page as an input to the item's calculation. supported marks a tied item (with calculation when derived), contradicted a draft value that differs from the source, partially_supported an item of which only some figures tie, and insufficient_evidence with no quotations an item no source states. A figure rounded from the source, such as $83.9 million from 83,872 in thousands, ties without a calculation. The viewer shows the draft with numbered tick marks beside the highlighted source passage and exports the ticked draft. comparison.template may name a template in place of comparison.fields; CiteLLM expands it to that template's fields and returns them with the template name. comparison.versions marks a comparison of two versions of one document: the first document is the earlier version and the second the later one, each with a label. Each field then carries change: changed, added (stated only in the later version), removed (stated only in the earlier one) or same, and a term a version does not contain is insufficient_evidence with no quotations. Validation reports a change that contradicts the evidence, such as added with a quotation from the earlier version, or same with different quoted wording and no change_note. The viewer shows both versions side by side with the changed words marked. consistency marks a review of one PDF as a consistency check, and check on each claim describes it. A match check is one figure stated in two or more places, each quoted as its own passage (A, B, C), with check.figures listing the figure as printed in each; a total check is components in calculation that should add up to the printed total in check.total. supported marks figures that agree or add up, contradicted marks a difference, and insufficient_evidence marks a figure stated in only one place. The server recomputes each check from the quoted figures and returns a validation error when the status disagrees, or marks the check not conclusive when units or rounding leave it unclear. The viewer lists the checks by group and shows the places side by side.

citellm_review_evidence

Saves a named list of comparison fields to the signed-in CiteLLM account as a template, or deletes a saved template, by template_id. Saving an existing template_id replaces it. Each account keeps up to 50 templates of up to 30 fields.

citellm_save_template

Searches the text of an imported PDF. queries returns the total matches, page and surrounding text for up to 25 case-insensitive phrases; definitions returns the full text of up to 10 defined terms ("Term" means ...), with the pages it spans and the other defined terms it uses; list_definitions returns every defined term with the page that defines it. Results locate text and are not evidence by themselves.

citellm_search_pdf

Opens CiteLLM's embedded file picker, where the user selects one or more PDFs from their device, up to the review limit, and uploads them directly to CiteLLM. Use when the user asks for the picker or when the attachment cannot be transferred another way. Accepts an optional upload_id to reuse an upload ticket that was already prepared. Each uploaded PDF's session_id is available for page reading and reviews.

citellm_open_pdf_upload

Checks quotation locations and numeric and calculation requirements for a complete set of claims across one or more imported PDFs, without opening a viewer or saving a review. A later call can give the draft_id of an earlier one with changes: the claims to replace, add or remove by claim number, all claims being checked again. Returns ready_to_display, a draft_id for exactly the claims checked, quotation-location failures in evidence_errors and numeric or calculation failures in validation_errors, retaining valid matches. Numeric claims pass only when their values appear in matched quotations or a calculation recomputes from quoted inputs. Matching confirms where text appears, not whether a claim is true. comparison.fields lists the items extracted from each PDF, each with an id and label, for two or more documents. Each claim then sets field, with exactly one claim per field per document; its text is that document's value, and a value the PDF does not state is insufficient_evidence with no quotations. The viewer shows a fields-by-documents table whose cells open the highlighted source passage. requirements lists the requirements that the documents are checked against in a gap analysis, with a title and an id, short label and optional full text for each. Each claim then sets requirement, with exactly one claim per requirement across all documents: supported means covered, partially_supported means partly covered, and insufficient_evidence with no quotations means a gap. A partly covered or gap requirement also sets missing, one sentence on what the documents leave out. The viewer shows a requirements table with status counts and exports a gap report. tie_out ties a draft PDF to its sources, such as a covenant compliance certificate to the credit agreement and financial statements it reports on: session_id is the draft and tie_out.source_session_ids lists the source PDFs, in place of additional_documents. Each claim is then one figure or factual statement in the draft, in draft order, with draft_quote holding the exact draft text and source_session_id naming the source PDF its evidence quotes; a quotation from another source PDF names it in its own source_session_id, so one item can cite several sources, up to 12 quotations per item. A source_session_id may give only the first 8 characters of a session ID that the request gives in full. A session a quotation names that tie_out.source_session_ids leaves out is added as a source. A quotation may also come from the draft itself, under the draft's session_id, away from the text the item ticks: on another page, such as the draft's own schedule, or on the same page as an input to the item's calculation. supported marks a tied item (with calculation when derived), contradicted a draft value that differs from the source, partially_supported an item of which only some figures tie, and insufficient_evidence with no quotations an item no source states. A figure rounded from the source, such as $83.9 million from 83,872 in thousands, ties without a calculation. The viewer shows the draft with numbered tick marks beside the highlighted source passage and exports the ticked draft. comparison.template may name a template in place of comparison.fields; CiteLLM expands it to that template's fields and returns them with the template name. comparison.versions marks a comparison of two versions of one document: the first document is the earlier version and the second the later one, each with a label. Each field then carries change: changed, added (stated only in the later version), removed (stated only in the earlier one) or same, and a term a version does not contain is insufficient_evidence with no quotations. Validation reports a change that contradicts the evidence, such as added with a quotation from the earlier version, or same with different quoted wording and no change_note. The viewer shows both versions side by side with the changed words marked. consistency marks a review of one PDF as a consistency check, and check on each claim describes it. A match check is one figure stated in two or more places, each quoted as its own passage (A, B, C), with check.figures listing the figure as printed in each; a total check is components in calculation that should add up to the printed total in check.total. supported marks figures that agree or add up, contradicted marks a difference, and insufficient_evidence marks a figure stated in only one place. The server recomputes each check from the quoted figures and returns a validation error when the status disagrees, or marks the check not conclusive when units or rounding leave it unclear. The viewer lists the checks by group and shows the places side by side.

citellm_validate_evidence