get_clinical_study
ChatGPTGet one twin trial by its id (the id of a search_clinical_studies hit) or a registry id. The record holds the card fields (title, phase, status, dates, diseases, drugs, sponsors), and conditions (ICD-11, Orphanet), arms (drugs, roles, routes), orgs, registry_ids, records (the registry records, readable with twin_record), outcome (class, basis, evidence), endpoints, events and papers, with the twin release. Eligibility criteria, design and sites are not in the twin's tables: read the registry record with twin_record. Cite the trial by primary_registry_id and the release.
get_disease_targets
ChatGPTGenetics-anchored target module for a disease, from Open Targets (CC0). Grounds the Step 4 mechanism map in real associations instead of prose.
get_fda_guideline_full
ChatGPTRetrieve the full text of an FDA guidance document.
get_ich_guideline_full
ChatGPTRetrieve the full text of an ICH guideline from the database.
search_clinical_studies
ChatGPTSearch the twin's trials with structured filters. Enum values (any casing/spacing is normalised): phases: phase_1 | phase_2 | phase_3 | phase_4 | NA (also "Phase 3", "PHASE3", "3") statuses: recruiting | completed | terminated | withdrawn | active_not_recruiting | not_yet_recruiting | suspended | unknown icd11_chapters are ICD-11 chapter numbers as strings ("07"); rare_disease_only keeps trials with an Orphanet-coded condition. search_text matches titles, conditions, drug names, sponsors and registry ids. For anything these filters cannot say, use twin_sql. Returns {"ceus": [trial], "total", "page", "page_size", "total_pages"}. Each trial carries its UUID id (for get_clinical_study), primary_registry_id, title, phase, status, dates, diseases, drugs, lead_sponsors, countries and the outcome class. For staff callers trials also carry clinical_trial_prediction (signal0_prob_cal / signal1_prob_cal) when the phase is scored. The reply's release names the twin release read.
search_fda_guidelines
ChatGPTSemantic search over FDA guidance documents. Table-of-contents entries and bare headings are left out, so every chunk returned is guidance text.
search_ich_guidelines
ChatGPTSemantic search over ICH regulatory guidelines. Use this to find relevant ICH guidance (e.g. E6 GCP, M4 CTD structure, Q8 pharmaceutical development) for a specific question.
twin_catalog
ChatGPTStart here. What the MondoTwin twin serves you: its domains and active releases, and its tables with their columns, row counts and purposes. table: a word table names must contain, "" for all (past a size budget the columns are left out: ask for those tables by name). MondoTwin's twin of the pharma world: trials, drugs and products, companies, deals, papers and documents, compiled by the twin engine into versioned releases and served read only. Each domain you are served (pharma, drugs, and twin, the schema v1 model) is one active release, read through the views of <domain>_<principal>; twin_catalog says which. Conventions of the v1 tables (where a table lacks a column, it does not follow it yet): - ids are named the same in every table (trial_id, drug_id, org_id, paper_id, document_id, disease_id, deal_id); find one with twin_find (any name, or an outside id such as nct:NCT01234567, pmid:..., unii:..., cik:...). - sources (text[]) on every row: the record keys <source>:<id> it was built from; twin_record(key) shows the raw record. Cite a fact by its record keys (and passage_id for text), with the release id from twin_catalog. - known_from: the earliest date the row's sources were public. For an "as of <date>" question add known_from <= '<date>'. - coverage: which source was read, for which territory or filer, through which date. Before answering "none", "not approved" or "no deal", check coverage says that source was read for that scope; otherwise say "unknown". - Text (labels, product information, filings, abstracts) is in passage, searchable with twin_text; markings on a document or passage name its licence. The content is licensed for internal research: quote only what you need.
twin_find
ChatGPTFind the twin's ids for a name or an outside id, through entity_name and external_id (where served). text: any name (a brand, an INN, a development code, a sponsor string, a former company name: exact case-insensitive matches first, then names whose words start with yours), or an outside id as scheme:value (nct:NCT01234567, pmid:12345678, unii:..., cik:0000012345, lei:...) or a bare value. limit: 1 to 100. Returns {"matches": [{"entity_type", "entity_id", "via" (external_id, name, name prefix), "matched", "kind", "schema"}]}. Then query the entity's tables with twin_sql by that id (drug_id, trial_id, org_id, ...).
twin_record
ChatGPTRead the raw source record a row was built from, as captured: the registry entry, the label, the filing, the paper. key: a record key <source>:<id> from a row's sources, such as clinicaltrials.gov:NCT01234567 (an old key reaches its record through the source's aliases). Served only from a source capture the principal is cleared for. offset: where to start reading, for a record longer than one page (20,000 characters). Returns {"key", "record_key" (the key it is stored under), "snapshot", "version" (the capture), "chars", "offset", "text" (the record's JSON, one page), "next_offset" (null at the end)}.
twin_sql
ChatGPTRun one read-only PostgreSQL query (SELECT, WITH, VALUES or TABLE) over the twin's views; the served schemas are on the search path, so write trial, not twin_shadow.trial. Read only, 15 s at most, no function that holds locks, sleeps or reads outside the twin. max_rows: 1 to 500 (default 100); the reply also stops at about 60 KB, and a cell over 2,000 characters is cut: select the columns you need, and use LIMIT, WHERE, left(text, n) or jsonb paths for wide values. Returns {"columns": [...], "rows": [[...], ...], "row_count", "more" (rows were left out), "elapsed_ms"} and "note" when the reply was cut. Find ids with twin_find first; check twin_catalog for the tables and columns. Before saying a thing does not exist, query coverage (where served) for whether the source was read; for an "as of" question filter on known_from; cite the sources record keys of the rows you used.
twin_text
ChatGPTSearch the text of documents (registry records, labels, EU product information, filings, abstracts) by words, with web-search syntax: words are ANDed, "quoted phrase", OR, -excluded. limit: 1 to 25. Returns {"passages": [{"passage_id", "document_id", "title", "kind", "section", "snippet" (the matching words in <b>), "text" (its first 2,000 characters), "chars" (its length), "markings" (its licence), "links" (the drugs, trials and companies the document concerns), "rank"}]}, best first. Read a whole passage with twin_sql (SELECT text FROM passage WHERE passage_id = ...). Cite the passage_id and the document's title.