@juspay/neurolink 9.67.2 → 9.67.3

package/dist/providers/openaiChatCompletionsBase.js

@@ -0,0 +1,643 @@
+/**
+ * Abstract base class for providers that talk to an OpenAI chat-completions
+ * shaped HTTP endpoint. Owns the entire request/stream/tool-loop pipeline
+ * so concrete providers only declare configuration + provider-specific
+ * quirks (env var names, default model, error mapping).
+ *
+ * Currently extended by:
+ *   - OpenAICompatibleProvider (generic /v1/chat/completions backend)
+ *   - LiteLLMProvider          (LiteLLM proxy server)
+ *   - DeepSeekProvider         (api.deepseek.com)
+ *
+ * Subclasses provide:
+ *   - getProviderName() / getDefaultModel() / formatProviderError() (abstract)
+ *   - optional overrides: getFallbackModelName, getFallbackModels,
+ *     adjustBuildBodyOptions, onStreamStart, getAvailableModels
+ *
+ * Nothing here imports from "ai" or "@ai-sdk/*". The base class is a
+ * direct HTTP client + multi-step tool-execution loop driven by SSE.
+ */
+import { BaseProvider } from "../core/baseProvider.js";
+import { DEFAULT_MAX_STEPS } from "../core/constants.js";
+import { streamAnalyticsCollector } from "../core/streamAnalytics.js";
+import { createProxyFetch } from "../proxy/proxyFetch.js";
+import { logger } from "../utils/logger.js";
+import { NoOutputGeneratedError } from "../utils/generationErrors.js";
+import { buildNoOutputSentinel, stampNoOutputSpan, } from "../utils/noOutputSentinel.js";
+import { composeAbortSignals, createTimeoutController, mergeAbortSignals, } from "../utils/timeout.js";
+import { emitToolEndFromStepFinish } from "../utils/toolEndEmitter.js";
+import { resolveToolChoice } from "../utils/toolChoice.js";
+import { transformToolExecutions } from "../utils/transformationUtils.js";
+import { buildAPIError, buildBody, buildToolsForOpenAI, createChunkQueue, createDeferredAnalytics, mapNeuroLinkToolChoice, mergeUsage, messageBuilderToOpenAI, parseSSEStream, stringifyToolOutput, stripTrailingSlash, v3ResponseFormatToOpenAI, v3ToolChoiceToOpenAI, v3ToolsToOpenAI, } from "./openaiChatCompletionsClient.js";
+/**
+ * Abstract HTTP+SSE provider for OpenAI chat-completions-shaped endpoints.
+ */
+export class OpenAIChatCompletionsProvider extends BaseProvider {
+    config;
+    resolvedModel;
+    constructor(providerName, modelName, sdk, config) {
+        super(modelName, providerName, sdk);
+        this.config = config;
+    }
+    // ===========================================================================
+    // Optional overridable hooks
+    // ===========================================================================
+    /**
+     * Model name to return when `getDefaultModel()` is empty AND
+     * auto-discovery via `/models` finds nothing. Default "gpt-3.5-turbo".
+     */
+    getFallbackModelName() {
+        return "gpt-3.5-turbo";
+    }
+    /**
+     * Hardcoded model names returned from `getAvailableModels()` when the
+     * remote `/models` endpoint can't be reached. Default empty.
+     */
+    getFallbackModels() {
+        return [];
+    }
+    /**
+     * Hook to mutate the `buildBody` options before the wire body is
+     * constructed. Default identity. Override for model-specific quirks
+     * (e.g. LiteLLM's Gemini 2.5 maxTokens skip).
+     */
+    adjustBuildBodyOptions(_modelId, opts) {
+        return opts;
+    }
+    /**
+     * Hook called once at the start of every `executeStream` invocation.
+     * Return lifecycle listeners (onUsage / onFinish) to receive deferred
+     * analytics events as the stream progresses. Default returns undefined
+     * (no extra wiring). LiteLLM uses this for the OTel span wrap with cost.
+     */
+    onStreamStart(_modelId) {
+        return undefined;
+    }
+    /**
+     * Returns true if `resolveModelName` should fall back to fetching
+     * `getAvailableModels()` and picking the first one when no explicit
+     * model is configured. Default true. Subclasses with a non-empty
+     * `getDefaultModel()` will never hit this branch anyway.
+     */
+    shouldAutoDiscoverModel() {
+        return true;
+    }
+    // ===========================================================================
+    // Public/protected concrete methods (shared by all subclasses)
+    // ===========================================================================
+    supportsTools() {
+        return true;
+    }
+    /**
+     * Returns a minimal V3-shaped model used by BaseProvider's `generate()`
+     * non-streaming path. Driven by the parent's `generateText`. The
+     * streaming path bypasses this entirely.
+     */
+    async getAISDKModel() {
+        const modelId = await this.resolveModelName();
+        return this.buildDelegatingModel(modelId);
+    }
+    async resolveModelName() {
+        if (this.resolvedModel) {
+            return this.resolvedModel;
+        }
+        const explicit = this.modelName || this.getDefaultModel();
+        if (explicit && explicit.trim() !== "") {
+            this.resolvedModel = explicit;
+            if (this.modelName !== explicit) {
+                this.refreshHandlersForModel(explicit);
+            }
+            return explicit;
+        }
+        if (this.shouldAutoDiscoverModel()) {
+            try {
+                const available = await this.getAvailableModels();
+                if (available.length > 0) {
+                    this.resolvedModel = available[0];
+                    this.refreshHandlersForModel(available[0]);
+                    logger.info(`🔍 Auto-discovered model: ${available[0]} from ${available.length} available models`);
+                    return available[0];
+                }
+            }
+            catch (err) {
+                logger.warn("Model auto-discovery failed, using fallback:", err);
+            }
+        }
+        const fallback = this.getFallbackModelName();
+        this.resolvedModel = fallback;
+        this.refreshHandlersForModel(fallback);
+        return fallback;
+    }
+    buildDelegatingModel(modelId) {
+        const url = `${stripTrailingSlash(this.config.baseURL)}/chat/completions`;
+        const fetchImpl = createProxyFetch();
+        const apiKey = this.config.apiKey;
+        const providerName = this.providerName;
+        const adjustBuildBodyOptions = this.adjustBuildBodyOptions.bind(this);
+        const getTimeoutForOptions = (opts) => this.getTimeout((opts ?? {}));
+        return {
+            specificationVersion: "v3",
+            provider: providerName,
+            modelId,
+            supportedUrls: {},
+            doGenerate: async (options) => {
+                const messages = messageBuilderToOpenAI(options.prompt);
+                const body = buildBody({
+                    modelId,
+                    messages,
+                    options: adjustBuildBodyOptions(modelId, {
+                        maxTokens: options.maxOutputTokens,
+                        temperature: options.temperature,
+                        topP: options.topP,
+                        presencePenalty: options.presencePenalty,
+                        frequencyPenalty: options.frequencyPenalty,
+                        seed: options.seed,
+                        stopSequences: options.stopSequences,
+                    }),
+                    tools: v3ToolsToOpenAI(options.tools),
+                    ...(options.toolChoice
+                        ? { toolChoice: v3ToolChoiceToOpenAI(options.toolChoice) }
+                        : {}),
+                    streaming: false,
+                    ...(options.responseFormat
+                        ? {
+                            responseFormat: v3ResponseFormatToOpenAI(options.responseFormat),
+                        }
+                        : {}),
+                });
+                const timeoutController = createTimeoutController(getTimeoutForOptions(options), providerName, "generate");
+                const composedSignal = composeAbortSignals(options.abortSignal, timeoutController?.controller.signal);
+                let res;
+                try {
+                    res = await fetchImpl(url, {
+                        method: "POST",
+                        headers: {
+                            "Content-Type": "application/json",
+                            Authorization: `Bearer ${apiKey}`,
+                        },
+                        body: JSON.stringify(body),
+                        ...(composedSignal ? { signal: composedSignal } : {}),
+                    });
+                }
+                finally {
+                    timeoutController?.cleanup();
+                }
+                if (!res.ok) {
+                    throw await buildAPIError(url, body, res);
+                }
+                const json = (await res.json());
+                const choice = json.choices?.[0];
+                const text = (typeof choice?.message?.content === "string"
+                    ? choice.message.content
+                    : "") ?? "";
+                const content = [];
+                if (text.length > 0) {
+                    content.push({ type: "text", text });
+                }
+                for (const tc of choice?.message?.tool_calls ?? []) {
+                    content.push({
+                        type: "tool-call",
+                        toolCallId: tc.id,
+                        toolName: tc.function.name,
+                        input: tc.function.arguments ?? "",
+                    });
+                }
+                const rawFinish = choice?.finish_reason;
+                const unified = rawFinish === "length"
+                    ? "length"
+                    : rawFinish === "tool_calls" || rawFinish === "function_call"
+                        ? "tool-calls"
+                        : rawFinish === "content_filter"
+                            ? "content-filter"
+                            : "stop";
+                return {
+                    content,
+                    finishReason: { unified, raw: rawFinish ?? "stop" },
+                    usage: {
+                        inputTokens: {
+                            total: json.usage?.prompt_tokens,
+                            noCache: json.usage?.prompt_tokens,
+                            cacheRead: undefined,
+                            cacheWrite: undefined,
+                        },
+                        outputTokens: {
+                            total: json.usage?.completion_tokens,
+                            text: json.usage?.completion_tokens,
+                            reasoning: undefined,
+                        },
+                    },
+                    warnings: [],
+                    request: { body },
+                    response: {
+                        ...(json.id ? { id: json.id } : {}),
+                        ...(json.model ? { modelId: json.model } : {}),
+                        headers: {},
+                        body: json,
+                    },
+                };
+            },
+            doStream: () => {
+                throw new Error(`${providerName}: doStream is not implemented on the delegating model — the streaming path uses executeStream directly.`);
+            },
+        };
+    }
+    /**
+     * Streaming path — drives the chat-completions endpoint directly. No
+     * streamText, no AI SDK orchestrator. Tool calls, multi-step loops,
+     * telemetry, abort handling all inline.
+     */
+    async executeStream(options, _analysisSchema) {
+        this.validateStreamOptions(options);
+        const startTime = Date.now();
+        const timeout = this.getTimeout(options);
+        const timeoutController = createTimeoutController(timeout, this.providerName, "stream");
+        // Consumer-driven abort: fires when the async iterator is closed early
+        // (caller breaks out of `for await`). Without this the background
+        // `loopPromise` keeps reading SSE and running tools indefinitely.
+        const consumerAbortController = new AbortController();
+        const abortSignal = mergeAbortSignals([
+            options.abortSignal,
+            timeoutController?.controller.signal,
+            consumerAbortController.signal,
+        ]).signal;
+        let modelId;
+        let toolsRecord;
+        let openAITools;
+        let openAIToolChoice;
+        let conversation;
+        try {
+            modelId = await this.resolveModelName();
+            const shouldUseTools = !options.disableTools && this.supportsTools();
+            toolsRecord = shouldUseTools
+                ? options.tools || (await this.getAllTools())
+                : {};
+            openAITools = shouldUseTools
+                ? buildToolsForOpenAI(toolsRecord)
+                : undefined;
+            openAIToolChoice = mapNeuroLinkToolChoice(resolveToolChoice(options, toolsRecord, shouldUseTools));
+            const initialMessages = await this.buildMessagesForStream(options);
+            conversation = messageBuilderToOpenAI(initialMessages);
+        }
+        catch (setupErr) {
+            timeoutController?.cleanup();
+            throw setupErr;
+        }
+        const url = `${stripTrailingSlash(this.config.baseURL)}/chat/completions`;
+        const fetchImpl = createProxyFetch();
+        const maxSteps = options.maxSteps || DEFAULT_MAX_STEPS;
+        const emitter = this.neurolink?.getEventEmitter();
+        const toolsUsed = [];
+        const toolExecutionSummaries = [];
+        const { usagePromise, finishPromise, resolveUsage, resolveFinish } = createDeferredAnalytics();
+        const { pushChunk, nextChunk } = createChunkQueue();
+        // Per-provider lifecycle hook (e.g. OTel span wrap for LiteLLM).
+        const lifecycle = this.onStreamStart(modelId);
+        const loopPromise = this.runStreamLoop({
+            maxSteps,
+            modelId,
+            url,
+            apiKey: this.config.apiKey,
+            fetchImpl,
+            abortSignal,
+            options,
+            conversation,
+            openAITools,
+            openAIToolChoice,
+            toolsRecord,
+            emitter,
+            toolsUsed,
+            toolExecutionSummaries,
+            pushChunk,
+            resolveUsage,
+            resolveFinish,
+        });
+        // Closure-scoped capture: the runStreamLoop's catch block stashes the
+        // underlying provider error here so we can pass it through to
+        // buildNoOutputSentinel for richer telemetry (matches the pattern in
+        // openAI.ts / litellm.ts where onError preserves the upstream cause).
+        let capturedProviderError;
+        // Parameter named `error` so the compiled `capturedProviderError = error`
+        // assignment matches the regression-grep in test:context 6.14.
+        const captureProviderError = (error) => {
+            capturedProviderError = error;
+        };
+        if (lifecycle?.onUsage) {
+            usagePromise.then(lifecycle.onUsage).catch(() => {
+                // usage may never resolve if the stream is aborted before completion
+            });
+        }
+        if (lifecycle?.onFinish) {
+            finishPromise
+                .then((reason) => lifecycle.onFinish?.(reason, capturedProviderError))
+                .catch(() => {
+                /* swallowed by design — see above */
+            });
+        }
+        const providerName = this.providerName;
+        const transformedStream = async function* () {
+            let contentYielded = 0;
+            try {
+                for (;;) {
+                    const chunk = await nextChunk();
+                    if ("done" in chunk) {
+                        break;
+                    }
+                    if ("content" in chunk &&
+                        typeof chunk.content === "string" &&
+                        chunk.content.length > 0) {
+                        contentYielded++;
+                    }
+                    yield chunk;
+                }
+                // Surface any error that the loop threw after we drained the queue.
+                await loopPromise;
+                // No-output path: stream completed normally but yielded zero text.
+                // Build an enriched sentinel + stamp the active OTel span so
+                // Pipeline B (ContextEnricher) surfaces a WARNING-level Langfuse
+                // observation instead of silently succeeding.
+                if (contentYielded === 0 && toolsUsed.length === 0) {
+                    logger.warn(`${providerName}: Stream produced no output — emitting enriched sentinel`);
+                    const fauxNoOutput = new NoOutputGeneratedError({
+                        message: "Stream produced no output",
+                    });
+                    const sentinel = await buildNoOutputSentinel(fauxNoOutput, undefined, capturedProviderError);
+                    stampNoOutputSpan(sentinel);
+                    yield sentinel;
+                }
+            }
+            catch (streamError) {
+                if (NoOutputGeneratedError.isInstance(streamError)) {
+                    const sentinel = await buildNoOutputSentinel(streamError, undefined, capturedProviderError);
+                    stampNoOutputSpan(sentinel);
+                    yield sentinel;
+                    return;
+                }
+                const sentinel = await buildNoOutputSentinel(streamError, undefined, capturedProviderError);
+                stampNoOutputSpan(sentinel);
+                yield sentinel;
+                throw streamError;
+            }
+            finally {
+                if (!consumerAbortController.signal.aborted) {
+                    consumerAbortController.abort();
+                }
+            }
+        };
+        const result = {
+            stream: transformedStream(),
+            provider: this.providerName,
+            model: this.modelName,
+            analytics: streamAnalyticsCollector.createAnalytics(this.providerName, this.modelName, {
+                textStream: (async function* () { })(),
+                usage: usagePromise,
+                finishReason: finishPromise,
+            }, Date.now() - startTime, {
+                requestId: options.requestId ??
+                    `${this.providerName}-stream-${Date.now()}`,
+                streamingMode: true,
+            }),
+            toolsUsed,
+            metadata: {
+                startTime,
+                streamId: `${this.providerName}-${Date.now()}`,
+            },
+        };
+        // Lazy getter: every read transforms the live `toolExecutionSummaries`
+        // through the canonical `transformToolExecutions()` so consumers see
+        // `{name, input, output, duration}[]` (codebase convention), while still
+        // reflecting tools appended during streaming.
+        Object.defineProperty(result, "toolExecutions", {
+            enumerable: true,
+            configurable: true,
+            get: () => transformToolExecutions(toolExecutionSummaries.map((s) => ({
+                toolName: s.toolName,
+                input: s.input,
+                output: s.output,
+                duration: s.endTime.getTime() - s.startTime.getTime(),
+            }))),
+        });
+        loopPromise
+            .finally(() => timeoutController?.cleanup())
+            .catch((error) => {
+            captureProviderError(error);
+        });
+        return result;
+    }
+    async runStreamLoop(args) {
+        const { maxSteps, modelId, url, apiKey, fetchImpl, abortSignal, options, conversation, openAITools, openAIToolChoice, toolsRecord, emitter, toolsUsed, toolExecutionSummaries, pushChunk, resolveUsage, resolveFinish, } = args;
+        try {
+            let stepFinish = null;
+            let stepUsage;
+            for (let step = 0; step < maxSteps; step++) {
+                const stepResult = await this.streamOneStep({
+                    modelId,
+                    url,
+                    apiKey,
+                    fetchImpl,
+                    abortSignal,
+                    options,
+                    conversation,
+                    openAITools,
+                    openAIToolChoice,
+                    pushChunk,
+                });
+                stepFinish = stepResult.finishReason;
+                if (stepResult.usage) {
+                    stepUsage = mergeUsage(stepUsage, stepResult.usage);
+                }
+                if (stepResult.toolCalls.size === 0) {
+                    break;
+                }
+                await this.executeToolBatch({
+                    stepResult,
+                    conversation,
+                    toolsRecord,
+                    emitter,
+                    toolsUsed,
+                    toolExecutionSummaries,
+                    options,
+                });
+            }
+            resolveUsage({
+                promptTokens: stepUsage?.prompt_tokens ?? 0,
+                completionTokens: stepUsage?.completion_tokens ?? 0,
+                totalTokens: stepUsage?.total_tokens ?? 0,
+            });
+            resolveFinish(stepFinish ?? "stop");
+            pushChunk({ done: true });
+            return {
+                finishReason: stepFinish ?? "stop",
+                usage: stepUsage,
+            };
+        }
+        catch (err) {
+            logger.error(`${this.providerName}: Stream error`, {
+                error: err instanceof Error ? err.message : String(err),
+            });
+            resolveUsage({ promptTokens: 0, completionTokens: 0, totalTokens: 0 });
+            resolveFinish("error");
+            pushChunk({ done: true });
+            throw err;
+        }
+    }
+    async streamOneStep(args) {
+        const body = buildBody({
+            modelId: args.modelId,
+            messages: args.conversation,
+            options: this.adjustBuildBodyOptions(args.modelId, args.options),
+            tools: args.openAITools,
+            ...(args.openAIToolChoice !== undefined
+                ? { toolChoice: args.openAIToolChoice }
+                : {}),
+            streaming: true,
+        });
+        const res = await args.fetchImpl(args.url, {
+            method: "POST",
+            headers: {
+                "Content-Type": "application/json",
+                Authorization: `Bearer ${args.apiKey}`,
+            },
+            body: JSON.stringify(body),
+            ...(args.abortSignal ? { signal: args.abortSignal } : {}),
+        });
+        if (!res.ok) {
+            throw await buildAPIError(args.url, body, res);
+        }
+        if (!res.body) {
+            throw new Error(`${this.providerName}: stream response had no body`);
+        }
+        return parseSSEStream(res.body, (delta) => {
+            args.pushChunk({ content: delta });
+        });
+    }
+    async executeToolBatch(args) {
+        const { stepResult, conversation, toolsRecord, emitter, toolsUsed, toolExecutionSummaries, options, } = args;
+        const toolCallsForMessage = [];
+        for (const [, t] of stepResult.toolCalls) {
+            toolCallsForMessage.push({
+                id: t.id,
+                type: "function",
+                function: { name: t.name, arguments: t.argsBuffered },
+            });
+        }
+        conversation.push({
+            role: "assistant",
+            content: stepResult.text.length > 0 ? stepResult.text : null,
+            tool_calls: toolCallsForMessage,
+        });
+        for (const [, t] of stepResult.toolCalls) {
+            const startedAt = new Date();
+            let input;
+            try {
+                input = JSON.parse(t.argsBuffered || "{}");
+            }
+            catch {
+                input = t.argsBuffered;
+            }
+            let output;
+            let errorMsg;
+            const toolDef = toolsRecord[t.name];
+            emitter?.emit("tool:start", {
+                toolName: t.name,
+                toolCallId: t.id,
+                input,
+            });
+            if (!toolDef || typeof toolDef.execute !== "function") {
+                errorMsg = `Tool '${t.name}' is not registered.`;
+                output = { error: errorMsg };
+            }
+            else {
+                try {
+                    output = await toolDef.execute(input, {});
+                }
+                catch (err) {
+                    errorMsg = err instanceof Error ? err.message : String(err);
+                    output = { error: errorMsg };
+                }
+            }
+            const endedAt = new Date();
+            toolsUsed.push(t.name);
+            toolExecutionSummaries.push({
+                toolCallId: t.id,
+                toolName: t.name,
+                input,
+                output,
+                ...(errorMsg ? { error: errorMsg } : {}),
+                startTime: startedAt,
+                endTime: endedAt,
+            });
+            conversation.push({
+                role: "tool",
+                tool_call_id: t.id,
+                content: stringifyToolOutput(output),
+            });
+        }
+        const justExecuted = toolExecutionSummaries.slice(-stepResult.toolCalls.size);
+        emitToolEndFromStepFinish(emitter, justExecuted.map((s) => ({
+            toolName: s.toolName,
+            output: s.output,
+            ...(s.error ? { error: s.error } : {}),
+        })));
+        try {
+            await this.handleToolExecutionStorage(justExecuted.map((s) => ({
+                toolCallId: s.toolCallId,
+                toolName: s.toolName,
+                input: s.input,
+                output: s.output,
+            })), justExecuted.map((s) => ({
+                toolCallId: s.toolCallId,
+                toolName: s.toolName,
+                output: s.output,
+            })), options, new Date());
+        }
+        catch (err) {
+            logger.warn(`[${this.constructor.name}] Failed to store tool executions`, {
+                provider: this.providerName,
+                error: err instanceof Error ? err.message : String(err),
+            });
+        }
+    }
+    /**
+     * Default implementation hits `${baseURL}/models`. Subclasses with a
+     * different endpoint path, caching, or fallback strategy should override.
+     */
+    async getAvailableModels() {
+        try {
+            const modelsUrl = `${stripTrailingSlash(this.config.baseURL)}/models`;
+            logger.debug(`Fetching available models from: ${modelsUrl}`);
+            const proxyFetch = createProxyFetch();
+            const controller = new AbortController();
+            const t = setTimeout(() => controller.abort(), 5000);
+            const response = await proxyFetch(modelsUrl, {
+                headers: {
+                    Authorization: `Bearer ${this.config.apiKey}`,
+                    "Content-Type": "application/json",
+                },
+                signal: controller.signal,
+            });
+            clearTimeout(t);
+            if (!response.ok) {
+                logger.warn(`Models endpoint returned ${response.status}: ${response.statusText}`);
+                return this.getFallbackModels();
+            }
+            const data = await response.json();
+            if (!data.data || !Array.isArray(data.data)) {
+                logger.warn("Invalid models response format");
+                return this.getFallbackModels();
+            }
+            const models = data.data.map((model) => model.id).filter(Boolean);
+            if (logger.shouldLog("debug")) {
+                logger.debug(`Discovered ${models.length} models:`, models);
+            }
+            return models.length > 0 ? models : this.getFallbackModels();
+        }
+        catch (error) {
+            logger.warn(`[${this.constructor.name}] Failed to fetch models from endpoint:`, error);
+            return this.getFallbackModels();
+        }
+    }
+    async getFirstAvailableModel() {
+        const models = await this.getAvailableModels();
+        return models[0] || this.getFallbackModelName();
+    }
+}