> ## Documentation Index > Fetch the complete documentation index at: https://docs.naga.ac/llms.txt > Use this file to discover all available pages before exploring further. # Classify content > Classifies text and images against a moderation model's categories. The response carries an `id` the gateway mints, `mdr-` followed by 28 random characters; the `model` that judged; and `results[]` with one verdict per moderated item. This operation returns no `usage` block. Requires an API key. `model` names a model from Naga's catalog that serves `moderations`, and the gateway selects a provider for it. `input` takes one of three shapes: a single string, an array of strings, or an array of typed parts. A typed part is `{"type": "text", "text": …}` or `{"type": "image_url", "image_url": {"url": …}}`, and the two may be mixed. A part carrying any other `type` is refused, with both accepted types named. An array holds either all strings or all typed parts. An image `url` is an `http(s)` URL or a `data:` URI carrying the bytes inline. Any other scheme, including `file:`, is refused while the body is read. A remote URL is fetched before the model is called. A URL that cannot be read, or that exceeds the deployment's inline-media limit, is refused. `results[]` arrives in the order the moderation model returned it and is passed through unchanged, neither padded to the number of inputs nor trimmed to them. Each entry carries four fields. * `flagged` — whether the item tripped any category. This is the model's own verdict where it sent one, and "any category true" where it sent categories alone. * `categories` — the per-category booleans, in the model's own order. The key set belongs to that model rather than to a fixed list, so the map is read by key. * `category_scores` — the scores, keyed exactly as `categories`. * `category_applied_input_types` — which kinds of input each category was applied to. Present only when the model sends it, and otherwise absent rather than `null`. The account is charged on input tokens alone, at the model's published prices. There is no output side to charge. Moderation models report no usage, so the count is the gateway's own: text is tokenized here, and every `image_url` part costs a flat 250 input tokens regardless of the image. Every response below `200` carries an object under `error` with a machine-readable `type`, a human-readable `message`, and, where the failure concerns one field, `param`. `429` and `503` carry `Retry-After` with the seconds to wait. ## OpenAPI ````yaml /openapi/naga-api.json post /v1/moderations openapi: 3.1.0 info: title: NagaAI API description: >- Naga routes inference requests to a pool of upstream providers behind one endpoint. Requests are written in the OpenAI or the Anthropic dialect, and the gateway selects a provider that serves the requested model. Two credentials exist and they are not interchangeable. An API key reaches models. A provisioning key manages the account: its keys, its balance, and its usage. Each operation declares which one it accepts. Accounts are charged for the tokens a request consumes, at the published price of the model that served it. termsOfService: https://naga.ac/legal/terms contact: name: NagaAI url: https://naga.ac version: '2026-03-31' summary: Unified AI API with OpenAI and Anthropic compatibility layers. servers: - url: https://api.naga.ac description: Production - url: http://localhost:8500 description: Local development security: [] tags: - name: Chat Completions description: >- OpenAI-compatible chat completion requests, including streaming, tools, and multimodal input. - name: Messages description: >- Anthropic-compatible Messages API requests used by Claude-style clients and agents. - name: Responses description: >- OpenAI-compatible Responses API requests for structured input, tools, and streaming output items. - name: Embeddings description: >- Create vector embeddings for search, retrieval, clustering, and ranking use cases. - name: Images description: >- Generate or edit images with multimodal models and OpenAI-compatible payloads. - name: Audio description: >- Generate speech or convert uploaded audio into transcriptions and translations. - name: Moderations description: Classify text and image inputs for safety policy categories. - name: Models description: List available models, capabilities, and pricing metadata. - name: Startups description: >- Public metadata about AI startups and model providers surfaced by the gateway. - name: Account description: Provisioning-key protected account balance and usage endpoints. - name: API Keys description: Provisioning-key protected API key management endpoints. paths: /v1/moderations: post: tags: - Moderations summary: Classify content description: > Classifies text and images against a moderation model's categories. The response carries an `id` the gateway mints, `mdr-` followed by 28 random characters; the `model` that judged; and `results[]` with one verdict per moderated item. This operation returns no `usage` block. Requires an API key. `model` names a model from Naga's catalog that serves `moderations`, and the gateway selects a provider for it. `input` takes one of three shapes: a single string, an array of strings, or an array of typed parts. A typed part is `{"type": "text", "text": …}` or `{"type": "image_url", "image_url": {"url": …}}`, and the two may be mixed. A part carrying any other `type` is refused, with both accepted types named. An array holds either all strings or all typed parts. An image `url` is an `http(s)` URL or a `data:` URI carrying the bytes inline. Any other scheme, including `file:`, is refused while the body is read. A remote URL is fetched before the model is called. A URL that cannot be read, or that exceeds the deployment's inline-media limit, is refused. `results[]` arrives in the order the moderation model returned it and is passed through unchanged, neither padded to the number of inputs nor trimmed to them. Each entry carries four fields. * `flagged` — whether the item tripped any category. This is the model's own verdict where it sent one, and "any category true" where it sent categories alone. * `categories` — the per-category booleans, in the model's own order. The key set belongs to that model rather than to a fixed list, so the map is read by key. * `category_scores` — the scores, keyed exactly as `categories`. * `category_applied_input_types` — which kinds of input each category was applied to. Present only when the model sends it, and otherwise absent rather than `null`. The account is charged on input tokens alone, at the model's published prices. There is no output side to charge. Moderation models report no usage, so the count is the gateway's own: text is tokenized here, and every `image_url` part costs a flat 250 input tokens regardless of the image. Every response below `200` carries an object under `error` with a machine-readable `type`, a human-readable `message`, and, where the failure concerns one field, `param`. `429` and `503` carry `Retry-After` with the seconds to wait. operationId: create_moderation requestBody: content: application/json: schema: $ref: '#/components/schemas/Moderation' examples: basic: summary: Classify one string value: model: omni-moderation-latest input: I want to hurt them. required: true responses: '200': description: >- The verdict. `results` carries one entry per input, each with the flag, the per-category verdicts and the scores behind them. content: application/json: schema: $ref: '#/components/schemas/ModerationsResponse' examples: basic: summary: One result, flagged value: id: mdr-4tKq7Zc1Nb9Wx2Ry8Vd3Ms6Ph0Q model: omni-moderation-latest results: - flagged: true categories: violence: true harassment: false category_scores: violence: 0.9124 harassment: 0.0217 '400': description: |- The request is malformed: a field failed its constraint, the body is not valid JSON, or the model name is unknown. content: application/json: schema: $ref: '#/components/schemas/ErrorEnvelope' '401': description: >- The API key is missing, malformed, disabled, or does not belong to an account that may call this route. content: application/json: schema: $ref: '#/components/schemas/ErrorEnvelope' examples: invalid-api-key: summary: 401 — the key is not one this gateway serves value: error: type: invalid_api_key message: >- The provided API key is invalid. Please ensure your API key is correct and active. You can find more information about obtaining a key on our website: https://naga.ac If you just created this key, wait a few minutes and try again. '402': description: The account has run out of credits. Top up the balance and retry. content: application/json: schema: $ref: '#/components/schemas/ErrorEnvelope' examples: insufficient-credits: summary: 402 — the account has no credits left value: error: type: insufficient_quota message: >- More credits required to process this request. Visit https://naga.ac/dashboard/credits to add credits. '403': description: >- The account may not use this model, or the content was blocked by the moderation policy. content: application/json: schema: $ref: '#/components/schemas/ErrorEnvelope' '408': description: |- The request body was not received within the time the gateway waits for it. content: application/json: schema: $ref: '#/components/schemas/ErrorEnvelope' '410': description: The model is deprecated and no longer served. Pick a current one. content: application/json: schema: $ref: '#/components/schemas/ErrorEnvelope' '413': description: The request body exceeds the configured size limit. content: application/json: schema: $ref: '#/components/schemas/ErrorEnvelope' '422': description: |- The request was understood but its usage could not be metered, so it was not served. content: application/json: schema: $ref: '#/components/schemas/ErrorEnvelope' '429': description: >- The account exceeded its rate limit. `Retry-After` carries the wait in seconds. content: application/json: schema: $ref: '#/components/schemas/ErrorEnvelope' examples: rate-limited: summary: 429 — too many requests; `Retry-After` carries the wait value: error: type: rate_limit_exceeded message: Rate limit reached. Please try again in 3s. '500': description: >- The gateway failed for a reason of its own. Report it — this status is reserved for Naga's own bugs and is never an upstream's verdict. content: application/json: schema: $ref: '#/components/schemas/ErrorEnvelope' '503': description: |- No upstream is available right now, or the gateway is shedding load. `Retry-After` carries the suggested wait. content: application/json: schema: $ref: '#/components/schemas/ErrorEnvelope' security: - bearerAuth: [] - xApiKeyAuth: [] x-codeSamples: - lang: bash label: curl source: |- curl https://api.naga.ac/v1/moderations \ -H "Authorization: Bearer $NAGA_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "model": "omni-moderation-latest", "input": "I want to hurt them." }' - lang: python label: openai-python source: | import os from openai import OpenAI client = OpenAI( api_key=os.environ["NAGA_API_KEY"], base_url="https://api.naga.ac/v1", ) result = client.moderations.create( model="omni-moderation-latest", input="I want to hurt them.", ) print(result) - lang: javascript label: openai-node source: | import OpenAI from "openai"; const client = new OpenAI({ apiKey: process.env.NAGA_API_KEY, baseURL: "https://api.naga.ac/v1", }); const result = await client.moderations.create({ model: "omni-moderation-latest", input: "I want to hurt them.", }); console.log(result); components: schemas: Moderation: type: object description: The request body of `POST /v1/moderations`. required: - model - input properties: model: type: string description: >- The moderation model, named by its id or by one of its aliases. A name this gateway does not serve on this route is refused with 400 `model_not_found`; a name it served and has since retired, with 410 `model_deprecated`. input: $ref: '#/components/schemas/ModerationInput' description: >- What to moderate. It is billed on tokens counted here rather than reported by the model: text is tokenized locally, and each `image_url` part costs a flat 250 input tokens whatever the image is. ModerationsResponse: allOf: - type: object - type: object required: - id - model - results properties: id: type: string description: >- The identifier of this moderation, minted by the gateway: `mdr-` followed by 28 random letters and digits. It is Naga's own and matches no identifier the upstream may have used. model: type: string description: >- The model that moderated, as the canonical id the requested name resolved to — an alias sent in the request comes back resolved here. The value is this gateway's catalog name, not anything the provider reported about itself. results: type: array items: $ref: '#/components/schemas/ModerationResultResponse' description: >- The verdicts, in the order the moderation model returned them, normally one per input item. The array is passed through as it came: nothing here pads it out to the number of inputs or trims it to them. description: >- The body of a successful `POST /v1/moderations`. It carries no `usage` block, unlike the gateway's other non-LLM routes; the request is billed all the same, on tokens counted here rather than reported by the model. ErrorEnvelope: type: object description: |- `{"error": {…}}` — the body of every non-2xx answer outside `/v1/messages`. required: - error properties: error: $ref: '#/components/schemas/ErrorBody' description: >- The failure itself, and the envelope's only member. Every non-2xx HTTP status arrives with it, and a 2xx answer never does. ModerationInput: oneOf: - type: string title: Plain text - type: array title: List of texts items: type: string - type: array title: List of typed parts items: $ref: '#/components/schemas/ModerationPart' description: >- What to moderate: one string, an array of strings, or an array of typed parts — `text` and `image_url` objects, which may be mixed in one array. An `input` of `[]` is accepted by the route, though the union published here refuses it: both array arms match the empty array, and `oneOf` demands exactly one match. ModerationResultResponse: allOf: - type: object - type: object required: - flagged - categories - category_scores properties: flagged: type: boolean description: >- Whether this input tripped any category. The moderation model's own verdict when it sent one; when it sent `categories` without a verdict, the gateway derives this as "any category true". categories: type: object description: >- The per-category booleans the moderation model reported, in its own order. The key set is that model's. There is no fixed list of categories here, and two models need not agree on one, so a client reads the map rather than named keys. A value the model sends as something other than a boolean is read as `false`. additionalProperties: type: boolean propertyNames: type: string category_scores: type: object description: >- The per-category scores the moderation model reported, keyed exactly like `categories`. A score the model sends as something other than a number fails the whole request rather than arriving as a zero the caller cannot tell from a real one. additionalProperties: type: number format: double propertyNames: type: string category_applied_input_types: type: - object - 'null' description: >- Which kinds of input each category was applied to, as the moderation model labeled them. Sent only when the model sends it: the key is absent from the body otherwise, never `null`. additionalProperties: type: array items: type: string propertyNames: type: string description: One verdict — the moderation of a single input item. ErrorBody: type: object description: |- The `error` member: what went wrong, in the caller's terms. `code` and `param` are optional AND nullable, and both halves are load bearing: the boundary omits them, the LLM dialect writes them as `null`. A schema that only allowed one of the two would contradict a body the gateway actually sends. required: - type - message properties: type: type: string description: >- The machine-readable class of the failure, for example `invalid_request_error` or `server_error`. message: type: string description: The human-readable sentence shown to the caller. code: type: - string - 'null' description: The finer-grained code, when the failure has one. param: type: - string - 'null' description: The request field the failure is about, when it is about one. ModerationPart: oneOf: - oneOf: - $ref: '#/components/schemas/TextInput' title: Text part - oneOf: - $ref: '#/components/schemas/ImageUrlInput' title: Image URL part description: >- One element of the typed-parts arm: a `text` part or an `image_url` part, told apart by the `type` the element carries itself. TextInput: type: object description: A text part of a typed-parts `input` array. required: - type - text properties: type: $ref: '#/components/schemas/TextTag' description: >- Always the literal `text`, and required: it is what selects this part shape. A part whose `type` is outside `text` and `image_url` is refused with a message naming both. text: type: string description: The text to moderate. ImageUrlInput: type: object description: An image part of a typed-parts `input` array. required: - type - image_url properties: type: $ref: '#/components/schemas/ImageUrlTag' description: >- Always the literal `image_url`, and required: it is what selects this part shape. A part whose `type` is outside `text` and `image_url` is refused with a message naming both. image_url: $ref: '#/components/schemas/ModerationImageUrl' description: The image to moderate. TextTag: type: string description: The tag of a text part — the literal `text`, its only value. enum: - text ImageUrlTag: type: string description: The tag of an image part — the literal `image_url`, its only value. enum: - image_url ModerationImageUrl: type: object description: Where an `image_url` part's image comes from. required: - url properties: url: type: string description: >- The image itself: an `http(s)` URL, or a `data:` URI carrying the bytes inline. Any other scheme, `file:` among them, is refused while the body is being read. A remote URL is fetched before the moderation model is called. One whose body exceeds the deployment's inline-media limit, or that cannot be read at all, is refused. securitySchemes: bearerAuth: type: http scheme: bearer description: 'An account API key. Send it as `Authorization: Bearer `.' xApiKeyAuth: type: apiKey in: header name: x-api-key description: >- An account API key. Send it as `x-api-key: `, the header Anthropic-style clients use. ````