@pipedog/openai 0.0.0 → 0.3.0

package/dist/base_llm.js

@@ -0,0 +1,619 @@
+// Copyright (c) 2024-2026, Daily
+// Copyright (c) 2026, pipedog contributors
+// SPDX-License-Identifier: BSD-2-Clause
+//
+// Ported from pipecat: src/pipecat/services/openai/base_llm.py
+/**
+ * Base LLM service implementation for services that use the AsyncOpenAI client.
+ *
+ * Porting notes:
+ * - The client is `OpenAI` from the official `openai` JS SDK, which talks
+ *   through `fetch` and runs on every runtime. Subclasses for
+ *   OpenAI-compatible providers override `createClient()`.
+ * - The SDK's `NOT_GIVEN` sentinel is `undefined` (`OPENAI_NOT_GIVEN`), which
+ *   leaves a field out of the request body. The store-mode settings hold it
+ *   for the sampling fields that are unset, as upstream does.
+ * - Upstream builds the client with an httpx client carrying connection limits
+ *   (`DefaultAsyncHttpxClient(limits=connection_limits(...))`). `fetch` has no
+ *   such object, so the SDK's own fetch is used.
+ * - The JS SDK has no `extra_body` request option: it sends unknown params
+ *   as-is. `extra_body` (from `_mergeExtraBody()` or `Settings.extra`) is
+ *   merged into the request JSON just before sending, like the Python client.
+ * - `APIStatusError` / `APITimeoutError` are `APIError` (with `status`) and
+ *   `APIConnectionTimeoutError` in the JS SDK.
+ * - Deprecated APIs are not ported: `BaseOpenAILLMService.InputParams` and the
+ *   `model` / `params` constructor options.
+ */
+import { abortable, currentSignal, rethrowIfCancelled, TimeoutError, waitFor, } from "@pipedog/asyncio";
+import { LLMContextConversionError, } from "@pipedog/core/adapters/base_llm_adapter";
+import { OPENAI_NOT_GIVEN, } from "@pipedog/core/adapters/services/open_ai_adapter";
+import { FunctionCallFromLLM, LLMContextFrame, LLMFullResponseEndFrame, LLMFullResponseStartFrame, LLMTextFrame, } from "@pipedog/core/frames/frames";
+import { LLMTokenUsage } from "@pipedog/core/metrics/metrics";
+import { LLMService } from "@pipedog/core/services/llm_service";
+import { LLMSettings } from "@pipedog/core/services/settings";
+import { ErrorCategory } from "@pipedog/core/utils/errors";
+import { isTimeoutException } from "@pipedog/core/utils/http";
+import { logger } from "@pipedog/core/utils/logger";
+import { pyJsonLoads } from "@pipedog/core/utils/py_json";
+import { tracedLlm } from "@pipedog/core/utils/tracing/service_decorators";
+import { assertGiven, isGiven, NOT_GIVEN } from "@pipedog/core/utils/types";
+import OpenAI, { APIConnectionTimeoutError, APIError, BadRequestError, UnprocessableEntityError, } from "openai";
+function errorMessage(e) {
+    return e instanceof Error ? e.message : String(e);
+}
+/**
+ * Settings for BaseOpenAILLMService.
+ *
+ * - `maxCompletionTokens`: maximum completion tokens to generate.
+ *
+ * Upstream overrides the inherited sampling fields (`frequencyPenalty`,
+ * `presencePenalty`, `seed`, `temperature`, `topP`, `maxTokens`) to also
+ * accept the OpenAI SDK's sentinel, which the service stores there so those
+ * fields can be passed through unchanged to the client. The SDK's sentinel is
+ * `undefined` here, which TypeScript can't add to an inherited field's type;
+ * the service writes it with `setOpenAINotGiven()`.
+ */
+export class OpenAILLMSettings extends LLMSettings {
+    maxCompletionTokens = NOT_GIVEN;
+    constructor(options = {}) {
+        super(options);
+        this._assignOptions(options);
+    }
+}
+/** The settings fields that hold the OpenAI SDK's sentinel when unset. */
+const OPENAI_SENTINEL_FIELDS = [
+    "frequencyPenalty",
+    "presencePenalty",
+    "seed",
+    "temperature",
+    "topP",
+    "maxTokens",
+    "maxCompletionTokens",
+];
+/**
+ * Store the OpenAI SDK's "not given" sentinel (`OPENAI_NOT_GIVEN`) in the
+ * given settings fields, so they are left out of requests.
+ *
+ * Options objects can't carry it (an `undefined` option means "not given" to
+ * Pipecat), so services set it after building their default settings.
+ *
+ * @param settings Store-mode settings to update in place.
+ * @param fields The fields to set. Defaults to every OpenAI sampling field.
+ */
+export function setOpenAINotGiven(settings, fields = OPENAI_SENTINEL_FIELDS) {
+    const self = settings;
+    for (const name of fields) {
+        self[name] = OPENAI_NOT_GIVEN;
+    }
+}
+/**
+ * OpenAI models that predate strict JSON schema replies: gpt-3.5, gpt-4 and
+ * gpt-4-turbo, the first gpt-4o snapshot, chatgpt-4o and the o1 previews.
+ */
+export const OPENAI_MODEL_WITHOUT_RESPONSE_SCHEMA = /^(gpt-3\.5|gpt-4($|-)|gpt-4o-2024-05-13|chatgpt-4o|o1-mini|o1-preview)/;
+export class BaseOpenAILLMService extends LLMService {
+    static Settings = OpenAILLMSettings;
+    /**
+     * Set by services for providers other than OpenAI. Without a key the OpenAI
+     * client falls back to OPENAI_API_KEY and would send it to that provider.
+     */
+    static _requiresApiKey = false;
+    /**
+     * Whether this service's API supports the "developer" message role.
+     *
+     * OpenAI's native API supports it, but some OpenAI-compatible services
+     * (e.g. Cerebras) do not. Subclasses that don't support it should set this
+     * to `false`, which causes the adapter to convert "developer" messages to
+     * "user" messages before sending them to the API.
+     */
+    static supportsDeveloperRole = true;
+    /**
+     * Whether the API can enforce a response schema. OpenAI-compatible services
+     * whose API cannot should set this to `false`.
+     */
+    static supportsResponseSchema = true;
+    _serviceTier;
+    _retryTimeoutSecs;
+    _retryOnTimeout;
+    _fullModelName = "";
+    _client;
+    /**
+     * Initialize the BaseOpenAILLMService.
+     *
+     * @param options.apiKey OpenAI API key. If null, uses environment variable.
+     * @param options.baseUrl Custom base URL for OpenAI API. If null, uses default.
+     * @param options.organization OpenAI organization ID.
+     * @param options.project OpenAI project ID.
+     * @param options.defaultHeaders Additional HTTP headers to include in requests.
+     * @param options.serviceTier Service tier to use (e.g., "auto", "flex", "priority").
+     * @param options.settings Runtime-updatable settings.
+     * @param options.retryTimeoutSecs How long an inference may go without
+     *   producing output before it is abandoned and re-issued, when
+     *   `retryOnTimeout` is set. Defaults to 5.0 seconds.
+     * @param options.retryOnTimeout Whether to re-issue the request once if the
+     *   first attempt produces no output within `retryTimeoutSecs`. The retry
+     *   is unbounded.
+     *   Other options are passed to the parent LLMService.
+     * @throws Error If the service is for a provider other than OpenAI and
+     *   `apiKey` is empty (upstream: ValueError).
+     */
+    constructor(options = {}) {
+        const cls = new.target;
+        if (cls._requiresApiKey && !options.apiKey) {
+            throw new Error(`${cls.name} requires an api_key.`);
+        }
+        // 1. Initialize default_settings with hardcoded defaults
+        const defaultSettings = new cls.Settings({
+            model: "gpt-4.1",
+            systemInstruction: null,
+            topK: null,
+            filterIncompleteUserTurns: false,
+            userTurnCompletionConfig: null,
+            extra: {},
+        });
+        setOpenAINotGiven(defaultSettings);
+        // 2. Apply settings delta (canonical API, always wins)
+        if (options.settings !== undefined && options.settings !== null) {
+            defaultSettings.applyUpdate(options.settings);
+        }
+        super({ ...options, settings: defaultSettings });
+        this._serviceTier = options.serviceTier ?? null;
+        this._retryTimeoutSecs =
+            options.retryTimeoutSecs === undefined ? 5.0 : options.retryTimeoutSecs;
+        this._retryOnTimeout = options.retryOnTimeout === undefined ? false : options.retryOnTimeout;
+        this._client = this.createClient(options);
+        if (this._settings.systemInstruction) {
+            logger.debug(`${this}: Using system instruction: ${this._settings.systemInstruction}`);
+        }
+    }
+    /**
+     * Create an AsyncOpenAI client instance.
+     *
+     * @param options.apiKey OpenAI API key.
+     * @param options.baseUrl Custom base URL for the API.
+     * @param options.organization OpenAI organization ID.
+     * @param options.project OpenAI project ID.
+     * @param options.defaultHeaders Additional HTTP headers.
+     *   Other options are the service's constructor options, which subclasses
+     *   may read.
+     * @returns Configured AsyncOpenAI client instance.
+     */
+    createClient(options = {}) {
+        return new OpenAI({
+            apiKey: options.apiKey ?? undefined,
+            baseURL: options.baseUrl ?? undefined,
+            organization: options.organization ?? null,
+            project: options.project ?? null,
+            defaultHeaders: options.defaultHeaders ?? undefined,
+        });
+    }
+    /**
+     * Check if this service can generate processing metrics.
+     *
+     * @returns True, as OpenAI service supports metrics generation.
+     */
+    canGenerateMetrics() {
+        return true;
+    }
+    /**
+     * Set the full AI model name.
+     *
+     * @param fullModelName The full name of the AI model to use.
+     */
+    setFullModelName(fullModelName) {
+        this._fullModelName = fullModelName;
+    }
+    /**
+     * Get the current full model name.
+     *
+     * @returns The full name of the AI model being used.
+     */
+    getFullModelName() {
+        return this._fullModelName;
+    }
+    /** Whether this service's API supports the "developer" message role. */
+    get supportsDeveloperRole() {
+        return this.constructor.supportsDeveloperRole;
+    }
+    /**
+     * Turn request parameters into the JSON body to send: `extra_body` is
+     * merged in over the other fields, as the Python client does.
+     */
+    _requestBody(params) {
+        const { extra_body: extraBody, ...rest } = params;
+        return { ...rest, ...(extraBody ?? {}) };
+    }
+    /**
+     * Get streaming chat completions from OpenAI API with optional timeout and retry.
+     *
+     * @param context Context to use for the chat completion. Contains messages,
+     *   tools, and tool choice.
+     * @returns Async stream of chat completion chunks.
+     */
+    async getChatCompletions(context) {
+        const adapter = this.getLlmAdapter();
+        this._logLlmResponse(context);
+        const paramsFromContext = await adapter.getLlmInvocationParams(context, {
+            systemInstruction: assertGiven(this._settings.systemInstruction),
+            convertDeveloperToUser: !this.supportsDeveloperRole,
+        });
+        const params = this.buildChatCompletionParams(paramsFromContext);
+        const create = () => abortable(this._client.chat.completions.create(this._requestBody(params), {
+            signal: currentSignal(),
+        }));
+        if (this._retryOnTimeout) {
+            try {
+                const chunks = await waitFor(() => create(), this._retryTimeoutSecs);
+                return chunks;
+            }
+            catch (e) {
+                if (e instanceof TimeoutError || e instanceof APIConnectionTimeoutError) {
+                    // Retry, this time without a timeout so we get a response
+                    logger.debug(`${this}: Retrying chat completion due to timeout`);
+                    const chunks = await create();
+                    return chunks;
+                }
+                throw e;
+            }
+        }
+        else {
+            const chunks = await create();
+            return chunks;
+        }
+    }
+    /**
+     * Build parameters for chat completion request.
+     *
+     * Subclasses can override this to customize parameters for different
+     * providers.
+     *
+     * @param paramsFromContext Parameters, derived from the LLM context, to use
+     *   for the chat completion. Contains messages, tools, and tool choice.
+     * @returns Dictionary of parameters for the chat completion request.
+     */
+    buildChatCompletionParams(paramsFromContext) {
+        const params = {
+            model: this._settings.model,
+            stream: true,
+            stream_options: { include_usage: true },
+            frequency_penalty: this._settings.frequencyPenalty,
+            presence_penalty: this._settings.presencePenalty,
+            seed: this._settings.seed,
+            temperature: this._settings.temperature,
+            top_p: this._settings.topP,
+            max_tokens: this._settings.maxTokens,
+            max_completion_tokens: this._settings.maxCompletionTokens,
+            service_tier: this._serviceTier !== null ? this._serviceTier : OPENAI_NOT_GIVEN,
+        };
+        // Messages, tools, tool_choice
+        Object.assign(params, paramsFromContext);
+        Object.assign(params, this._settings.extra);
+        return params;
+    }
+    /**
+     * Add provider-specific fields to a request's `extra_body`.
+     *
+     * The OpenAI client rejects keyword arguments it doesn't know, so fields
+     * specific to an OpenAI-compatible provider travel in `extra_body`, which
+     * the client merges into the request JSON as-is.
+     *
+     * @param params Request parameters, updated in place. An `extra_body`
+     *   already there, supplied through `Settings.extra`, wins key by key,
+     *   matching how `extra` overrides every other request parameter. It is
+     *   copied rather than modified.
+     * @param fields Fields to add. Unset and null values are omitted.
+     */
+    static _mergeExtraBody(params, fields) {
+        const extraBody = {};
+        for (const [name, value] of Object.entries(fields)) {
+            if (isGiven(value) && value !== null && value !== undefined) {
+                extraBody[name] = value;
+            }
+        }
+        Object.assign(extraBody, params.extra_body || {});
+        if (Object.keys(extraBody).length > 0) {
+            params.extra_body = extraBody;
+        }
+    }
+    /**
+     * Whether a model can enforce a response schema.
+     *
+     * OpenAI models before gpt-4o-mini and gpt-4o-2024-08-06 cannot.
+     *
+     * @param model The model name.
+     */
+    static modelSupportsResponseSchema(model) {
+        return !OPENAI_MODEL_WITHOUT_RESPONSE_SCHEMA.test(model);
+    }
+    /**
+     * Run a one-shot, out-of-band (i.e. out-of-pipeline) inference with the
+     * given LLM context.
+     *
+     * @param context The LLM context containing conversation history.
+     * @param options.maxTokens Optional maximum number of tokens to generate. If
+     *   provided, overrides the service's default max_tokens /
+     *   max_completion_tokens setting.
+     * @param options.systemInstruction Optional system instruction to use for
+     *   this inference. If provided, overrides any system instruction in the
+     *   context.
+     * @param options.responseSchema Optional JSON schema the reply must follow.
+     *   The service asks the provider to enforce it, so the reply is JSON text
+     *   matching the schema.
+     * @returns The LLM's response as a string, or null if no response is
+     *   generated.
+     */
+    async runInference(context, options = {}) {
+        const maxTokens = options.maxTokens ?? null;
+        const effectiveInstruction = options.systemInstruction || assertGiven(this._settings.systemInstruction);
+        const adapter = this.getLlmAdapter();
+        const invocationParams = await adapter.getLlmInvocationParams(context, {
+            systemInstruction: effectiveInstruction,
+            convertDeveloperToUser: !this.supportsDeveloperRole,
+        });
+        // Build params using the same method as streaming completions
+        const params = this.buildChatCompletionParams(invocationParams);
+        // Override for non-streaming
+        params.stream = false;
+        delete params.stream_options;
+        // Override max_tokens if provided
+        if (maxTokens !== null) {
+            // Use max_completion_tokens for newer models, fallback to max_tokens
+            if ("max_completion_tokens" in params) {
+                params.max_completion_tokens = maxTokens;
+            }
+            else {
+                params.max_tokens = maxTokens;
+            }
+        }
+        const responseSchema = this._checkResponseSchema(options.responseSchema);
+        if (responseSchema !== null) {
+            params.response_format = {
+                type: "json_schema",
+                json_schema: { name: "response", schema: responseSchema, strict: true },
+            };
+        }
+        // LLM completion
+        const response = (await abortable(this._client.chat.completions.create(this._requestBody(params), {
+            signal: currentSignal(),
+        })));
+        return response.choices[0]?.message.content ?? null;
+    }
+    /**
+     * Convert a chat completion's usage into Pipecat's token usage.
+     *
+     * Follows OpenAI's convention, where `completion_tokens` includes the
+     * reasoning tokens reported in `completion_tokens_details`. Providers that
+     * report reasoning elsewhere, or apart from `completion_tokens`, override
+     * this.
+     *
+     * @param usage The usage reported with the completion.
+     * @returns The token usage to report.
+     */
+    _tokenUsage(usage) {
+        const promptDetails = usage.prompt_tokens_details;
+        const cachedTokens = promptDetails ? (promptDetails.cached_tokens ?? null) : null;
+        // Tokens written into the prompt cache, billed above the input rate.
+        // Providers without prompt caching omit the field, which reads as "not
+        // reported" rather than zero.
+        const cacheWriteTokens = promptDetails
+            ? (promptDetails.cache_write_tokens ?? null)
+            : null;
+        const reasoningTokens = usage.completion_tokens_details
+            ? (usage.completion_tokens_details.reasoning_tokens ?? null)
+            : null;
+        return new LLMTokenUsage({
+            promptTokens: usage.prompt_tokens,
+            completionTokens: usage.completion_tokens,
+            totalTokens: usage.total_tokens,
+            cacheReadInputTokens: cachedTokens,
+            cacheCreationInputTokens: cacheWriteTokens,
+            reasoningTokens: reasoningTokens,
+        });
+    }
+    async _processContext(context) {
+        const functionsList = [];
+        const argumentsList = [];
+        const toolIdList = [];
+        let funcIdx = 0;
+        let functionName = "";
+        let args = "";
+        let toolCallId = "";
+        await this.startTtfbMetrics();
+        // Generate chat completions from LLMContext
+        const chunkStream = await this.getChatCompletions(context);
+        // Ensure the stream and its async iterator are closed on
+        // cancellation/exception to release the HTTP connection. Each chunk is
+        // awaited as a cancellation point, since the SDK's iterator is a foreign
+        // await.
+        const chunkIter = chunkStream[Symbol.asyncIterator]();
+        // Providers differ in how often they send usage: some once at the end,
+        // others a cumulative snapshot on every chunk. Holding the latest and
+        // reporting it after the stream keeps that to one report per completion.
+        let tokenUsage = null;
+        try {
+            try {
+                while (true) {
+                    const next = await abortable(chunkIter.next());
+                    if (next.done)
+                        break;
+                    const chunk = next.value;
+                    if (chunk.usage) {
+                        tokenUsage = this._tokenUsage(chunk.usage);
+                    }
+                    if (chunk.model && this.getFullModelName() !== chunk.model) {
+                        this.setFullModelName(chunk.model);
+                    }
+                    if (chunk.choices === null || chunk.choices === undefined || chunk.choices.length === 0) {
+                        continue;
+                    }
+                    await this.stopTtfbMetrics();
+                    const delta = chunk.choices[0]?.delta;
+                    if (!delta) {
+                        continue;
+                    }
+                    if (delta.tool_calls && delta.tool_calls.length > 0) {
+                        // A turn that only calls tools produces no answer text, so the
+                        // call itself is what the caller gets and TTFAT ends here rather
+                        // than going unmeasured.
+                        await this.stopTtfatMetrics();
+                        // We're streaming the LLM response to enable the fastest response
+                        // times. For text, we just yield each chunk as we receive it and
+                        // count on consumers to do whatever coalescing they need (eg. to
+                        // pass full sentences to TTS)
+                        //
+                        // If the LLM is a function call, we'll do some coalescing here. If
+                        // the response contains a function name, we'll yield a frame to
+                        // tell consumers that they can start preparing to call the
+                        // function with that name. We accumulate all the arguments for the
+                        // rest of the streamed response, then when the response is done,
+                        // we package up all the arguments and the function name and yield
+                        // a frame containing the function name and the arguments.
+                        const toolCall = delta.tool_calls[0];
+                        if (toolCall.index !== funcIdx) {
+                            functionsList.push(functionName);
+                            argumentsList.push(args || "{}");
+                            toolIdList.push(toolCallId);
+                            functionName = "";
+                            args = "";
+                            toolCallId = "";
+                            funcIdx += 1;
+                        }
+                        if (toolCall.function?.name) {
+                            functionName += toolCall.function.name;
+                            toolCallId = toolCall.id;
+                        }
+                        if (toolCall.function?.arguments) {
+                            // Keep iterating through the response to collect all the argument fragments
+                            args += toolCall.function.arguments;
+                        }
+                    }
+                    else if (delta.content) {
+                        await this._pushLlmText(delta.content);
+                    }
+                    // When gpt-4o-audio / gpt-4o-mini-audio is used for llm or stt+llm
+                    // we need to get LLMTextFrame for the transcript
+                    else if (delta.audio?.transcript) {
+                        await this.pushFrame(new LLMTextFrame({ text: delta.audio.transcript }));
+                    }
+                }
+            }
+            finally {
+                // Close the iterator first, then abort the request to release HTTP
+                // resources.
+                try {
+                    await chunkIter.return?.();
+                }
+                catch (e) {
+                    rethrowIfCancelled(e);
+                }
+                chunkStream.controller?.abort();
+            }
+        }
+        finally {
+            // Report even if the response is interrupted or cancelled mid-stream.
+            if (tokenUsage) {
+                await this.startLlmUsageMetrics(tokenUsage);
+            }
+        }
+        // if we got a function name and arguments, check to see if it's a function
+        // with a registered handler. If so, run the registered callback, save the
+        // result to the context, and re-prompt to get a chat answer. If we don't
+        // have a registered handler, raise an exception.
+        if (functionName) {
+            // added to the list as last function name and arguments not added to the list
+            functionsList.push(functionName);
+            argumentsList.push(args || "{}");
+            toolIdList.push(toolCallId);
+            const functionCalls = [];
+            for (let i = 0; i < functionsList.length; i++) {
+                const name = functionsList[i];
+                const rawArguments = argumentsList[i];
+                const toolId = toolIdList[i];
+                let parsedArguments;
+                try {
+                    parsedArguments = pyJsonLoads(rawArguments);
+                }
+                catch (e) {
+                    rethrowIfCancelled(e);
+                    logger.warning(`${this}: Failed to parse function call arguments: ${rawArguments}`);
+                    continue;
+                }
+                functionCalls.push(new FunctionCallFromLLM({
+                    context,
+                    toolCallId: toolId,
+                    functionName: name,
+                    arguments: parsedArguments,
+                }));
+            }
+            await this.runFunctionCalls(functionCalls);
+        }
+    }
+    /**
+     * Process frames for LLM completion requests.
+     *
+     * Handles LLMContextFrame to trigger LLM completions.
+     *
+     * @param frame The frame to process.
+     * @param direction The direction of frame processing.
+     */
+    async processFrame(frame, direction) {
+        await super.processFrame(frame, direction);
+        if (frame instanceof LLMContextFrame) {
+            try {
+                await this.pushFrame(new LLMFullResponseStartFrame());
+                await this.startProcessingMetrics();
+                await this._processContext(frame.context);
+            }
+            catch (e) {
+                rethrowIfCancelled(e);
+                if (isTimeoutException(e)) {
+                    await this._callEventHandler("on_completion_timeout");
+                    await this.pushError("LLM completion timeout", { exception: e });
+                }
+                else if (e instanceof LLMContextConversionError) {
+                    await this.pushError(e.message, { exception: e });
+                    // A conversion failure (e.g. an unsupported file MIME type, corrupt
+                    // base64 data) can't reach the API at all, but is just as much
+                    // evidence of an invalid file as a rejection from OpenAI itself, so
+                    // it gets the same best-effort cleanup.
+                    frame.context.removeInvalidFileMessage();
+                }
+                else {
+                    // Only the payload-shaped errors (bad request, unprocessable
+                    // content, payload too large — the latter has no dedicated OpenAI
+                    // exception class, so match its status code) are grounds to remove
+                    // a pending file message on a best-effort basis. The rest of the
+                    // 4xx range — auth, permissions, not-found, rate limiting — says
+                    // nothing about whether our request (or its file) was bad, and
+                    // removing the file there would discard it for no benefit: the file
+                    // isn't what needs fixing before the next retry can succeed. When a
+                    // message is removed as a result of this error, the fault lay in
+                    // application-supplied content and the context is repaired, so the
+                    // error is pushed as APPLICATION instead of letting the rejection
+                    // classify as permanent and cost the service its usability.
+                    const removed = (e instanceof BadRequestError ||
+                        e instanceof UnprocessableEntityError ||
+                        (e instanceof APIError && e.status === 413)) &&
+                        frame.context.removeInvalidFileMessage();
+                    await this.pushError(`Error during completion: ${errorMessage(e)}`, {
+                        exception: e,
+                        category: removed ? ErrorCategory.APPLICATION : null,
+                    });
+                }
+            }
+            finally {
+                await this.stopProcessingMetrics();
+                await this.pushFrame(new LLMFullResponseEndFrame());
+            }
+        }
+        else {
+            await this.pushFrame(frame, direction);
+        }
+    }
+}
+// @traced_llm on `_process_context`.
+{
+    const proto = BaseOpenAILLMService.prototype;
+    proto._processContext = tracedLlm(proto._processContext);
+}
+//# sourceMappingURL=base_llm.js.map