> ## Documentation Index > Fetch the complete documentation index at: https://docs.slng.ai/llms.txt > Use this file to discover all available pages before exploring further. # Create a chat completion > Route an OpenAI Chat Completions request. An agent ID and session ID are required, as headers (`X-Slng-Agent-Id`, `X-Slng-Session-Id`) or body fields (`slng_agent_id`, `slng_session_id`); the body field wins if both are sent. ## OpenAPI ````yaml /api-reference/context-router/context-router.oas.yaml post /chat/completions openapi: 3.1.0 info: title: Context Router API version: 1.0.0 summary: >- OpenAI-compatible chat endpoints with routing, response caching and per-call templating. description: > Routes OpenAI-compatible requests to a configured model and returns the standard OpenAI response. Two endpoints share the same auth, routing, caching and templating: * `POST /chat/completions` - Chat Completions protocol. Recommended. * `POST /responses` - Responses protocol. SLNG request extensions: `template_variables`, `slng_config`, `slng_agent_id`, `slng_session_id`, `slng_analytics`, `slng_pure_proxy`. Response headers: `x-slng-request-id`, `x-slng-response-source`, `x-slng-model`. servers: - url: https://{region}.context-router.slng.ai/v1 description: Regional base URL. Pick the region closest to your callers. variables: region: default: india enum: - india - us - indonesia description: Deployment region. Other regions may exist. security: - SlngApiKey: [] tags: - name: Chat description: OpenAI Chat Completions surface. - name: Responses description: OpenAI Responses-compatible endpoint. paths: /chat/completions: post: tags: - Chat summary: Create a chat completion description: > Route an OpenAI Chat Completions request. An agent ID and session ID are required, as headers (`X-Slng-Agent-Id`, `X-Slng-Session-Id`) or body fields (`slng_agent_id`, `slng_session_id`); the body field wins if both are sent. operationId: createChatCompletion parameters: - $ref: '#/components/parameters/XSlngAgentId' - $ref: '#/components/parameters/XSlngSessionId' - $ref: '#/components/parameters/XSlngRequestId' requestBody: required: true content: application/json: schema: $ref: '#/components/schemas/ChatCompletionRequest' examples: minimal: summary: Minimal routed call value: model: slng/auto messages: - role: system content: You are a helpful clinic assistant. - role: user content: What are your opening hours? templated: summary: Per-call personalization via template_variables value: model: slng/auto messages: - role: system content: >- You are Sarah, a scheduling assistant for {{clinic_name}}. - role: user content: hi, can you confirm my appointment? temperature: 0.3 max_tokens: 256 template_variables: clinic_name: Sunrise Clinic streaming: summary: Streaming with the usage trailer value: model: slng/auto messages: - role: user content: Hello stream: true stream_options: include_usage: true inlineConfig: summary: Bring your own model configuration inline value: model: slng/auto messages: - role: user content: Hello slng_config: tiers: '1': - model: gpt-4o-mini weight: 100 endpoint: url: https://api.openai.com/v1 api_key: sk-... responses: '200': description: >- Chat completion in OpenAI format, live or cached. See `x-slng-response-source`. headers: x-slng-request-id: $ref: '#/components/headers/XSlngRequestIdResponse' x-slng-response-source: $ref: '#/components/headers/XSlngResponseSource' x-slng-model: $ref: '#/components/headers/XSlngModel' content: application/json: schema: $ref: '#/components/schemas/ChatCompletionResponse' text/event-stream: schema: type: string description: > OpenAI SSE stream when `stream: true`, ending with `data: [DONE]`. Set `stream_options.include_usage` for a final `usage` chunk. '400': description: > Bad request: missing or malformed agent or session ID (`missing_client_id`, `invalid_client_id`), unknown pinned `model`, or invalid `slng_config`. headers: x-slng-request-id: $ref: '#/components/headers/XSlngRequestIdResponse' content: application/json: schema: $ref: '#/components/schemas/ErrorResponse' examples: missingClientId: value: error: message: Missing required agent id type: invalid_request_error code: missing_client_id '422': description: | Unprocessable request: a template variable had no value (`missing_template_variables`), `slng_config` was not a JSON object (`invalid_slng_config`), or a template variable exceeded its limits. headers: x-slng-request-id: $ref: '#/components/headers/XSlngRequestIdResponse' content: application/json: schema: $ref: '#/components/schemas/ErrorResponse' examples: missingTemplateVariables: value: error: message: >- Missing required template variables: appointment_time, city type: invalid_request_error param: template_variables code: missing_template_variables '429': $ref: '#/components/responses/UpstreamError' '502': $ref: '#/components/responses/UpstreamError' '504': $ref: '#/components/responses/UpstreamError' components: parameters: XSlngAgentId: name: X-Slng-Agent-Id in: header required: false description: > Stable agent ID. Scopes the cache — change it when the system prompt changes. Required unless sent as `slng_agent_id` (body wins). Max 256 chars; no whitespace, commas, pipes or braces. schema: $ref: '#/components/schemas/SlngId' example: clinic-scheduler-v1 XSlngSessionId: name: X-Slng-Session-Id in: header required: false description: | ID for a single call, constant across its turns. Required unless sent as `slng_session_id` (body wins). Same character rules as the agent ID. schema: $ref: '#/components/schemas/SlngId' example: call-8f2c1a4e-9b07-4d3a-a1e2-7c5f0d6b2a11 XSlngRequestId: name: X-Slng-Request-Id in: header required: false description: > Optional correlation ID, echoed back in `x-slng-request-id`. `X-Request-Id` is also accepted. schema: type: string schemas: ChatCompletionRequest: type: object required: - messages additionalProperties: true description: > OpenAI Chat Completions body plus SLNG extensions. Unlisted OpenAI fields pass through. properties: model: type: string default: slng/auto description: > `slng/auto` (or omit) routes across your configured models with failover. Pass an entry's `model` label to pin one — disabling failover and giving it a separate cache. Any other value returns a 400. examples: - slng/auto - gpt-4o-mini messages: type: array minItems: 1 description: > OpenAI messages array. `template_variables` substitution touches the system message only. items: $ref: '#/components/schemas/ChatMessage' temperature: type: number description: OpenAI parameter, passed through. top_p: type: number description: OpenAI parameter, passed through. max_tokens: type: integer description: OpenAI parameter, passed through. max_completion_tokens: type: integer description: OpenAI parameter, passed through. stream: type: boolean default: false description: > `true` for SSE streaming, `false` for a single JSON response. Supported for every model. stream_options: type: object additionalProperties: true description: Streaming options. properties: include_usage: type: boolean default: false description: | Set `true` while streaming to get the final usage chunk. tools: type: array description: > OpenAI tool definitions. Answers with tool calls, and the turn after a tool result, are not cached. items: type: object additionalProperties: true tool_choice: description: OpenAI tool choice, passed through. oneOf: - type: string - type: object additionalProperties: true template_variables: $ref: '#/components/schemas/TemplateVariables' slng_config: $ref: '#/components/schemas/SlngConfig' slng_agent_id: allOf: - $ref: '#/components/schemas/SlngId' description: > Body form of `X-Slng-Agent-Id`. Required unless the header is sent; wins if both. slng_session_id: allOf: - $ref: '#/components/schemas/SlngId' description: > Body form of `X-Slng-Session-Id`. Required unless the header is sent; wins if both. slng_analytics: type: boolean default: false description: > Set `true` to record the request and response (in-region) for analytics. Requires both IDs. slng_pure_proxy: type: boolean default: false description: > Set `true` to return the provider response with no caching or edits. Routing and failover still apply. ChatCompletionResponse: type: object additionalProperties: true description: > OpenAI Chat Completions response, live or cached. `usage` reports the cached answer's tokens on a hit, billed tokens on a live answer. properties: id: type: string object: type: string example: chat.completion created: type: integer model: type: string choices: type: array items: type: object additionalProperties: true properties: index: type: integer message: $ref: '#/components/schemas/ChatMessage' finish_reason: type: string usage: type: object additionalProperties: true properties: prompt_tokens: type: integer completion_tokens: type: integer total_tokens: type: integer ErrorResponse: type: object required: - error properties: error: type: object properties: message: type: string type: type: string example: invalid_request_error param: type: - string - 'null' code: type: string enum: - missing_client_id - invalid_client_id - missing_template_variables - invalid_slng_config description: > Error code. `slng_config` failures other than a non-object body return a 400 with a message beginning `invalid slng_config:`. SlngId: type: string maxLength: 256 pattern: ^[^\s,|{}]+$ description: | Opaque ID. Max 256 characters; no whitespace, commas, pipes or braces. ChatMessage: type: object required: - role additionalProperties: true properties: role: type: string enum: - system - user - assistant - tool content: description: Message content. String, or the standard OpenAI content-part array. oneOf: - type: string - type: array items: type: object additionalProperties: true - type: 'null' name: type: string tool_calls: type: array items: type: object additionalProperties: true tool_call_id: type: string TemplateVariables: type: object maxProperties: 64 description: > SLNG extension. `{name: value}` map substituted into `{{name}}` placeholders in the **system message only**. Names are letters, digits and underscore. Limits (over any returns 422): 64 variables, 64-char names, 4000-char values. A `{{name}}` with no value returns 422 (`missing_template_variables`); unused variables are ignored. Use for personalization that is spoken or echoed. Values that steer the answer (language, plan, region) belong in the prompt — steering here can make the cache serve another caller's answer. propertyNames: pattern: ^[A-Za-z0-9_]{1,64}$ additionalProperties: type: string maxLength: 4000 example: caller_name: Rajesh city: Mumbai SlngConfig: type: object required: - tiers additionalProperties: false description: > SLNG extension. Model endpoints and provider keys sent inline, replacing your saved configuration for that request. Treat as a credential (HTTPS, keep out of logs); under 256 KB serialized. Invalid config returns 400 (`invalid slng_config:`); a non-object returns 422 (`invalid_slng_config`). properties: tiers: type: object description: > Up to three tiers (`"1"`, `"2"`, `"3"`) tried in order. Within a tier, traffic splits by `weight` (summing to 100). A failed model (5xx, timeout, 429) retries once against the next option. propertyNames: enum: - '1' - '2' - '3' additionalProperties: type: array minItems: 1 items: $ref: '#/components/schemas/SlngConfigEntry' SlngConfigEntry: type: object required: - model - weight - endpoint additionalProperties: false properties: model: type: string description: > Label for this entry. Pass it as the request `model` to target it directly. weight: type: integer minimum: 1 maximum: 100 description: Share of the tier's traffic; weights in a tier sum to 100. endpoint: $ref: '#/components/schemas/SlngEndpoint' SlngEndpoint: type: object additionalProperties: false description: > Where the model is served, with its credentials. Required fields depend on `provider`; fields for another provider are rejected. properties: provider: type: string default: openai-compat enum: - openai-compat - openai-responses - azure - vertex - bedrock description: > `openai-compat` (default) for any OpenAI Chat Completions service (OpenAI, Groq, vLLM, and so on). `openai-responses` calls the upstream Responses API instead — for models that only work that way, typically reasoning models that refuse tools. url: type: string format: uri description: > Required for `openai-compat`, `openai-responses` and `azure`. For `azure`, the resource **root** (`https://my-resource.cognitiveservices.azure.com/`), not the deployment URL. api_key: type: string description: > Required for `openai-compat`, `openai-responses` and `azure`. Your provider key, not your SLNG key. auth_header: type: string description: > Optional (`openai-compat`). Custom key header, for example `x-goog-api-key` for Gemini. model_id: type: string description: > Model name sent to the provider when it differs from the `model` label. Required for `bedrock`. azure_deployment: type: string description: 'Required for `provider: azure`. The Azure deployment name.' api_version: type: string description: 'Required for `provider: azure`, for example `2024-12-01-preview`.' vertex_credentials: type: object additionalProperties: true description: > Required for `provider: vertex`. GCP service-account key as a JSON object. vertex_location: type: string description: 'Required for `provider: vertex`. The Vertex AI location.' vertex_project: type: string description: > Optional for `provider: vertex`. Defaults to the service-account key's project. aws_access_key_id: type: string description: 'Required for `provider: bedrock`.' aws_secret_access_key: type: string description: 'Required for `provider: bedrock`.' aws_region: type: string description: 'Required for `provider: bedrock`.' aws_session_token: type: string description: 'Optional for `provider: bedrock`, for temporary credentials.' examples: - url: https://api.openai.com/v1 api_key: sk-... - provider: azure url: https://my-resource.cognitiveservices.azure.com/ api_key: ... azure_deployment: gpt-4o-deploy api_version: 2024-12-01-preview - provider: bedrock aws_access_key_id: AKIA... aws_secret_access_key: ... aws_region: ap-south-1 model_id: anthropic.claude-3-5-sonnet-20241022-v2:0 headers: XSlngRequestIdResponse: description: > The router's ID for this request, on every response. Quote it in support requests. schema: type: string XSlngResponseSource: description: > Answer source: `llm` (live model) or `cache`. `router` and `analytics` appear only with those features. schema: type: string enum: - llm - cache - router - analytics XSlngModel: description: > The model that produced the answer, after any failover. Live answers only. schema: type: string responses: UpstreamError: description: > Upstream failure (provider `429`, `502` or `504`). Only `x-slng-request-id` is returned. headers: x-slng-request-id: $ref: '#/components/headers/XSlngRequestIdResponse' content: application/json: schema: $ref: '#/components/schemas/ErrorResponse' securitySchemes: SlngApiKey: type: http scheme: bearer description: > Send your SLNG API key as a Bearer token (an OpenAI SDK's `api_key` option). Provider credentials in `slng_config` authenticate to those providers. ````