We just launched four new task specific models, available on our Trustedrouter page. They are fine tuned to be very fast and very cheap at individual tasks, including: document extraction, grounded document Q&A, conversation summary, and classification routing. Below are instructions for using them.
1. Document Structured Extraction --Published
Endpoint: neurometric/document-structured-extraction
Use this to extract selected fields from document text. The caller supplies the document and an output schema defining the requested fields and types. Missing fields are returned as
null.
Example input:
{”document”:”Invoice 1042. Vendor: Acme Corporation. Date: 2026-09-01. Total: USD 250.00.”,”output_schema”:
{”invoice_id”:”string”,”vendor”:”string”,”date”:”string”,”total”:”number”,”currency”:”string”,”purchase_order”:”string|null”}}
Expected output:
{”invoice_id”:”1042”,”vendor”:”Acme Corporation”,”date”:”2026-09-01”,”total”:250,”currency”:”USD”,”purchase_order”:null}
It currently expects extracted text rather than a raw PDF.
2. Grounded Document QA
Endpoint: neurometric/grounded-document-qa
Use this to answer questions using only supplied document chunks. The response includes the answer and supporting source IDs. If the documents do not contain the answer, the endpoint abstains.
Example input:
{”chunks”:[{”id”:”POLICY-1”,”status”:”authoritative”,”text”:”Customers may return unused products within 30 days of purchase with a receipt.”},{”id”:”POLICY-
2”,”status”:”authoritative”,”text”:”Refunds are issued to the original payment method.”}],”question”:”How long do customers have to return an unused product?”}
Expected output:
{”answer”:”30 days”,”citations”:[”POLICY-1”]}
Example when the answer is unavailable:
{”answer”:null,”citations”:[]}
3. Conversation Summary
Endpoint: neurometric/conversation-summary
Use this to extract the latest decision, current status, outstanding actions, owners, deadlines, and material risks from a business conversation. Completed and superseded information
is excluded.
Example input:
{”conversation_id”:”THREAD-001”,”messages”:[{”id”:”MSG-001”,”body”:”FINAL DECISION: Launch the new customer portal”},{”id”:”MSG-002”,”body”:”CURRENT STATUS: Waiting for security
approval”},{”id”:”MSG-003”,”body”:”OPEN ACTION: Send the approval packet. OWNER: Amina. DUE: 2026-09-10.”},{”id”:”MSG-004”,”body”:”MATERIAL RISK: Security review may delay the
launch”}]}
Expected output:
{”decision”:”Launch the new customer portal”,”current_status”:”Waiting for security approval”,”open_items”:[{”action”:”Send the approval packet”,”owner”:”Amina”,”due_date”:”2026-09-
10”}],”risks”:[”Security review may delay the launch”]}
4. Classification Router
Endpoint: neurometric/classification-router
Use this to classify requests using a taxonomy supplied at request time. The caller provides labels, their meanings, a fallback label, and the requests to classify.
Example input:
{”labels”:{”Billing”:”charges or refunds”,”Technical”:”errors or outages”,”Other”:”anything else”},”fallback_label”:”Other”,”requests”:[{”id”:”REQ-1”,”text”:”Please refund this
duplicate charge.”},{”id”:”REQ-2”,”text”:”The API returns a 502 error.”},{”id”:”REQ-3”,”text”:”What is your office address?”}]}
Expected output:
{”REQ-1”:”Billing”,”REQ-2”:”Technical”,”REQ-3”:”Other”}