Chart reads
How the agent reads a FHIR chart
A chat request becomes a tool call. The tool fetches FHIR resources from your backend and returns chart context for the model to use in its answer.
From patient search to a focused read
Start with “Find Maria Garcia,” then open the matching chart. Follow up with a question about a specific section or document. The model selects from these tools; it does not construct an arbitrary FHIR query.
- search_patients
- Find a patient by name. The name is passed as a structured search parameter, never spliced into a query string.
- show_patient_info
- Open one patient's chart across 23 sections, reporting when a window was capped rather than implying the chart is empty.
- read_chart_section
- Read one section with code and date filters, and follow references for authors, encounter, facility, or provenance. The tool builds the query; the model never does.
- read_document
- Read what a note actually says. Decodes an inline text body only — a scan or pointer-only attachment is reported as unread, never as empty.
A read result has limits
The reader reports capped result windows, unsupported filters, failed reference lookups, and documents it cannot decode. Those details matter when interpreting an answer.
For example, a vaccine-code search can return no matches even when the chart has a vaccine recorded as free text. The tool reports that distinction so the agent can read the section without the code filter. A partial or unmatched search is not evidence that the patient never received the vaccine.
Document reading currently supports inline text. A scanned PDF or an attachment URL is reported as unread. The model can still misinterpret the returned context; check the underlying records when accuracy matters.
Where chart context goes
In external model modes, your messages and the records the agent reads are sent to your configured provider. Reads do not pause for approval. The write approval card controls saving a change, not sending read context to the model.
Last EHR stores no chart database or chat transcripts of its own. It is alpha software, and the public demo uses synthetic patients. The zero-key local walkthrough makes no model request, but follows a fixed sequence rather than answering arbitrary questions.