> ## Documentation Index > Fetch the complete documentation index at: https://docs.valiancehealth.com/llms.txt > Use this file to discover all available pages before exploring further. # Create a chat completion > An OpenAI-compatible chat completion endpoint that redacts PHI before the model sees it and restores the real values in the answer. ```python client = OpenAI(base_url="https:///api/v1/llm", api_key="hpx_...") client.chat.completions.create(model="", messages=[...]) ``` Per request: your message content is redacted (Malaysian IC/MRN/passport/phone, names, emails, account numbers — clinical content such as ages, dates like "day 3 post-op", vitals and dosages is preserved), the placeholders go to the model, and the model's answer has your values substituted back. The mapping lives only in the request's memory and is never stored. Every response carries `redaction_receipt`, including `rehydration_complete` and `unresolved_placeholder_count` — if the model altered our placeholders, the receipt says so rather than returning residue silently. **Tool calling** is supported: send OpenAI-style `tools`, receive `tool_calls` with your values re-hydrated. Your `tool` role results are redacted on the way out like any other message, and so are the arguments of an assistant turn you send back. **Not supported** (refused, not silently dropped): streaming while re-hydrating (`stream` with `redact_only=false`), tools combined with `stream`, non-text content parts, forcing a specific tool via `tool_choice`, and the deprecated `functions`/`function_call` form. ## OpenAPI ````yaml /api/openapi-commercial.yaml post /api/v1/llm/chat/completions openapi: 3.0.3 info: title: Healthproximate API version: 1.0.0 description: >- DRG classification, OMOP vocabulary services and a PHI-safe LLM gateway. Generated by `manage.py export_commercial_openapi` from an explicit operation allowlist (`app/openapi_commercial.py`). An endpoint appears here because someone published it, never because it exists. contact: email: admin@valiancehealth.ai servers: - url: https://api.v2.healthproximate.com security: [] tags: - name: DRG description: MS-DRG classification from coded input, free clinical text, or FHIR. - name: Vocabulary description: OMOP concept search, coding, translation and hierarchy resolution. - name: Gateway description: >- PHI-safe LLM access. Text is redacted before any model sees it and re-hydrated in the answer, and every response carries a receipt. paths: /api/v1/llm/chat/completions: post: tags: - Gateway summary: Create a chat completion description: >- An OpenAI-compatible chat completion endpoint that redacts PHI before the model sees it and restores the real values in the answer. ```python client = OpenAI(base_url="https:///api/v1/llm", api_key="hpx_...") client.chat.completions.create(model="", messages=[...]) ``` Per request: your message content is redacted (Malaysian IC/MRN/passport/phone, names, emails, account numbers — clinical content such as ages, dates like "day 3 post-op", vitals and dosages is preserved), the placeholders go to the model, and the model's answer has your values substituted back. The mapping lives only in the request's memory and is never stored. Every response carries `redaction_receipt`, including `rehydration_complete` and `unresolved_placeholder_count` — if the model altered our placeholders, the receipt says so rather than returning residue silently. **Tool calling** is supported: send OpenAI-style `tools`, receive `tool_calls` with your values re-hydrated. Your `tool` role results are redacted on the way out like any other message, and so are the arguments of an assistant turn you send back. **Not supported** (refused, not silently dropped): streaming while re-hydrating (`stream` with `redact_only=false`), tools combined with `stream`, non-text content parts, forcing a specific tool via `tool_choice`, and the deprecated `functions`/`function_call` form. operationId: v1_llm_chat_completions_create requestBody: content: application/json: schema: $ref: '#/components/schemas/ChatCompletionRequestRequest' application/x-www-form-urlencoded: schema: $ref: '#/components/schemas/ChatCompletionRequestRequest' multipart/form-data: schema: $ref: '#/components/schemas/ChatCompletionRequestRequest' required: true responses: '200': description: chat.completion with a redaction receipt attached '400': description: Invalid request, or model not approved '401': description: Authentication required '429': description: Rate limit or spend limit reached '503': description: Redaction failed — request refused security: - gateway_bearer_auth: [] components: schemas: ChatCompletionRequestRequest: type: object properties: model: type: string minLength: 1 description: >- A model id from the registry. External keys may only use models approved for PHI traffic. messages: type: array items: $ref: '#/components/schemas/ChatMessageRequest' max_tokens: type: integer maximum: 32000 minimum: 1 default: 1024 temperature: type: number format: double nullable: true top_p: type: number format: double nullable: true stop: type: string nullable: true minLength: 1 stream: type: boolean default: false redact_only: type: boolean default: false description: >- Return the model's answer WITHOUT substituting your values back (placeholders remain). Required if you want streaming, and useful when the answer is going somewhere that should not hold PHI. tools: type: array items: type: object additionalProperties: {} description: >- OpenAI-style function tools. Forwarded to the model; returned tool_calls come back with your values re-hydrated. tool_choice: description: >- "auto" (default) or "none". Forcing a specific call ("required", or a named function) is refused rather than downgraded to "auto", so a forced call never appears to have been honoured when it was not. conversation_key: type: string description: >- Supply the same value across turns so a given person keeps the same placeholder; multi-turn clients resend history and renaming mid-conversation confuses the model. maxLength: 200 required: - messages - model ChatMessageRequest: type: object properties: role: $ref: '#/components/schemas/ChatMessageRoleEnum' content: type: string nullable: true maxLength: 100000 tool_calls: type: array items: type: object additionalProperties: {} tool_call_id: type: string minLength: 1 description: >- Required on a `tool` message: it matches the result to the call that asked for it. maxLength: 200 required: - role ChatMessageRoleEnum: enum: - system - user - assistant - tool type: string description: |- * `system` - system * `user` - user * `assistant` - assistant * `tool` - tool securitySchemes: gateway_bearer_auth: type: http scheme: bearer description: >- Your hpx_ API key as a Bearer token, so OpenAI-compatible SDKs work unmodified. X-API-Key and first-party JWTs are also accepted. ````