BURKI CUSTOM PROVIDER: PRIMARY LLM WORKSHEET Free planning resource | Prepared October 26, 2026 This worksheet is free. Model, voice, telephone and hosting costs remain separate. No key, endpoint, assistant, provider request or call was configured for this resource. Observed results begin NOT RUN. 1. IDENTIFY THE COMPONENT Business task: Workspace and authorised configuration owner: Assistant and intended publication version: Conversation engine: Standard pipeline / GPT Live / not checked What the custom language model should control: Standard pipeline: the selected primary LLM writes conversation replies. GPT Live: the separate business LLM does not replace the conversational voice engine. A Custom Provider selection alone does not change engines. This worksheet is for an HTTPS OpenAI-compatible Chat Completions endpoint. It is not a Custom WebSocket, Responses-only or realtime-audio setup. It is also not a fallback list or a Web request connected action. 2. APPROVED ENDPOINT INVENTORY - NO SECRETS Provider / endpoint owner: Documented HTTPS API base: Approved exact model identifier: Account owner and intended credential scope: Secure credential storage reference, never actual key: Organisation LLM billing/credential mode explicitly BYO, verified by: Account pricing / units / terms checked at: Who authorises any later provider evaluation and its spending: Required behaviour to verify with the actual endpoint: Chat Completions request compatibility: Streaming response chunks: Generation options supported: Complete response and usage evidence: Tool-call/result behaviour, only if separately required: Failure and uncertain-response handling: A key entered in the assistant does not switch the account's LLM mode. Native Custom Provider admission requires explicit BYO and an HTTPS base. A compatibility label does not establish acceptance of every option or tool. Leave the actual running worker revision UNKNOWN until separately established. 3. CURRENT UI PATH FOR AN AUTHORISED BUILD Assistants -> New assistant -> Set up manually -> Receptionist Voice -> Model and provider choices -> Default AI stack AI provider -> Custom Provider AI model: approved exact identifier Provider API key: authorised credential entered through the actual product Base URL: provider's documented HTTPS API base New manual assistants currently default to GPT Live. Review the Conversation engine deliberately. For an intended Standard pipeline, review the separate recognition and speech providers as well. Do not change a saved assistant just to inspect these fields. Read Publish -> Published version before saving an existing assistant. Published assistant: saved edits remain drafts until Publish saved changes. Legacy live settings warning: a save changes connected calls. New creation, publication and telephone connection are separate stages. 4. FICTIONAL HARBOR SIGN DESK INVENTORY Purpose: answer approved public collection-counter questions Intended engine: Standard pipeline AI provider: Custom Provider AI model: REPLACE_WITH_APPROVED_MODEL_ID Base URL illustration: https://llm.example.com/v1 The example.com address is reserved illustrative material, not a live service. Credential: supplied securely by authorised owner, never written in this file LLM mode: explicit BYO, not assumed from the entered key Connected actions for this first task: none Provider acceptance / voice result: NOT TESTED ILLUSTRATIVE APPROVED FACTS: The public collection counter opens Tuesday and Thursday, 10:00-15:00, in the business's stated timezone. Counter hours do not establish whether a particular sign is ready. Staff confirm order status and collection arrangements through the approved office contact process. Replace these with actual verified business facts, timezone and contact process. 5. COPYABLE TASK INSTRUCTIONS Answer collection-counter questions using the approved business facts. Keep public opening hours separate from the readiness of a particular order. If the caller corrects a day or time, use the latest request and read back the relevant detail. If a fact is missing, say what cannot be confirmed and explain the approved staff contact process. Do not claim an order is ready, a collection is reserved or staff were notified without an actual verified result. This instruction does not install a lookup, staff queue or identity check. 6. ORIGINAL RESPONSE FIXTURE - NOT RUN Caller: Can I collect my sign on Friday? Sorry, Thursday at two. Expected interpretation: Latest requested time: Thursday, 14:00, in verified business timezone Counter hours: within approved Thursday opening window Order readiness: UNKNOWN Collection reservation / notification: NOT ESTABLISHED Proposed answer, not generated model output: Thursday at 14:00 is within the collection counter's opening hours. I can't confirm whether your sign is ready; staff need to check the order through the office process. 7. ACCEPTANCE MATRIX - ALL OBSERVATIONS NOT RUN A. Plain public question Expected: accurate approved hours; no order-ready claim. Observed: NOT RUN Configuration / complete result or error / evidence: B. Friday corrected to Thursday Expected: latest Thursday14:00 preserved; Friday superseded. Observed: NOT RUN Evidence: C. Fact missing from approved information Expected: unknown remains unknown; actual approved staff process explained. Observed: NOT RUN Evidence: D. Key / HTTPS / credential-mode blocker Expected: preserve actual admission error; resolve rather than invent readiness. Observed: NOT RUN Evidence: E. Response stream incomplete or incompatible Expected: incomplete output stays a failure, not a successful conversation. Observed: NOT RUN Request / returned chunks / usage / error references, with no secrets: F. Correct text but no heard speech Expected: inspect the distinct voice path; text alone is not audible acceptance. Observed: NOT RUN Evidence: G. Supported connected action added later, if separately authorised Expected: inspect tool arguments, actual result and resulting answer separately. Observed: NOT RUN Evidence / inner business result / remaining unknowns: Keep the original attempt when a response is uncertain. Do not repeatedly create fresh sessions to conceal an unresolved response. This worksheet authorises none. 8. COST AND DECISION RECORD Provider usage units and actual account charges: Burki platform component: Recognition and spoken-output components: Telephone / hosting / optional service components: Human review effort, actually measured: Failed attempts and unresolved cases retained: Scope accepted by / date / evidence: Exact running worker revision, if established: Provider compatibility accepted for this model/options: Actual caller-heard result: Remaining unknowns and next owner: Companion guide: https://burki.dev/blog/burki-custom-primary-llm-setup https://burki.dev/pricing