@librechat/agents 3.1.77-dev.1 → 3.1.77

package/dist/cjs/llm/openai/index.cjs

@@ -53,6 +53,47 @@ function normalizeHeaders(headers) {
     });
     return Object.fromEntries(output.entries());
 }
+function createUsageMetadata(usage) {
+    const usageMetadata = {
+        input_tokens: usage?.prompt_tokens ?? 0,
+        output_tokens: usage?.completion_tokens ?? 0,
+        total_tokens: usage?.total_tokens ?? 0,
+    };
+    if (usage == null) {
+        return usageMetadata;
+    }
+    const inputTokenDetails = {};
+    const outputTokenDetails = {};
+    let hasInputTokenDetails = false;
+    let hasOutputTokenDetails = false;
+    const audioInputTokens = usage.prompt_tokens_details?.audio_tokens;
+    const cachedInputTokens = usage.prompt_tokens_details?.cached_tokens;
+    const audioOutputTokens = usage.completion_tokens_details?.audio_tokens;
+    const reasoningOutputTokens = usage.completion_tokens_details?.reasoning_tokens;
+    if (audioInputTokens != null) {
+        inputTokenDetails.audio = audioInputTokens;
+        hasInputTokenDetails = true;
+    }
+    if (cachedInputTokens != null) {
+        inputTokenDetails.cache_read = cachedInputTokens;
+        hasInputTokenDetails = true;
+    }
+    if (audioOutputTokens != null) {
+        outputTokenDetails.audio = audioOutputTokens;
+        hasOutputTokenDetails = true;
+    }
+    if (reasoningOutputTokens != null) {
+        outputTokenDetails.reasoning = reasoningOutputTokens;
+        hasOutputTokenDetails = true;
+    }
+    if (hasInputTokenDetails) {
+        usageMetadata.input_token_details = inputTokenDetails;
+    }
+    if (hasOutputTokenDetails) {
+        usageMetadata.output_token_details = outputTokenDetails;
+    }
+    return usageMetadata;
+}
 function getExposedOpenAIClient(completions, responses, preferResponses) {
     const responsesClient = responses.client;
     if (responsesClient?.abortHandler != null) {
@@ -793,6 +834,54 @@ class ChatDeepSeek extends deepseek.ChatDeepSeek {
     static lc_name() {
         return 'LibreChatDeepSeek';
     }
+    _convertDeepSeekMessages(messages) {
+        return index._convertMessagesToOpenAIParams(messages, this.model, {
+            includeReasoningContent: true,
+        });
+    }
+    async _generate(messages$1, options, runManager) {
+        options.signal?.throwIfAborted();
+        const params = this.invocationParams(options);
+        if (params.stream === true) {
+            return super._generate(messages$1, options, runManager);
+        }
+        const messagesMapped = this._convertDeepSeekMessages(messages$1);
+        const response = await this.completionWithRetry({
+            ...params,
+            stream: false,
+            messages: messagesMapped,
+        }, {
+            signal: options.signal,
+            ...options.options,
+        });
+        const usageMetadata = createUsageMetadata(response.usage);
+        const generations = response.choices.map((part) => {
+            const text = part.message.content ?? '';
+            const generation = {
+                text,
+                message: this._convertCompletionsMessageToBaseMessage(part.message, response),
+            };
+            generation.generationInfo = {
+                finish_reason: part.finish_reason,
+                ...(part.logprobs != null ? { logprobs: part.logprobs } : {}),
+            };
+            if (messages.isAIMessage(generation.message)) {
+                generation.message.usage_metadata = usageMetadata;
+            }
+            generation.message = new messages.AIMessage(Object.fromEntries(Object.entries(generation.message).filter(([key]) => !key.startsWith('lc_'))));
+            return generation;
+        });
+        return {
+            generations,
+            llmOutput: {
+                tokenUsage: {
+                    promptTokens: usageMetadata.input_tokens,
+                    completionTokens: usageMetadata.output_tokens,
+                    totalTokens: usageMetadata.total_tokens,
+                },
+            },
+        };
+    }
     _getClientOptions(options) {
         if (!this.client) {
             const openAIEndpointConfig = {
@@ -817,7 +906,234 @@ class ChatDeepSeek extends deepseek.ChatDeepSeek {
         return requestOptions;
     }
     async *_streamResponseChunks(messages, options, runManager) {
-        yield* delayStreamChunks(super._streamResponseChunks(messages, options, runManager), this._lc_stream_delay);
+        yield* delayStreamChunks(this._streamResponseChunksWithReasoning(messages, options, runManager), this._lc_stream_delay);
+    }
+    /** Parses raw `<think>` fallback tags across chunks and emits sanitized DeepSeek stream chunks. */
+    async *_streamResponseChunksWithReasoning(messages$1, options, runManager) {
+        const stream = this._streamResponseChunksFromReasoningMessages(messages$1, options);
+        const thinkStartTag = '<think>';
+        const thinkEndTag = '</think>';
+        let tokensBuffer = '';
+        let isThinking = false;
+        for await (const chunk of stream) {
+            if (options.signal?.aborted === true) {
+                throw new Error('AbortError');
+            }
+            const reasoningContent = chunk.message.additional_kwargs.reasoning_content;
+            if (reasoningContent != null && reasoningContent !== '') {
+                yield* this._yieldDeepSeekStreamChunk(chunk, runManager);
+                continue;
+            }
+            const text = chunk.text;
+            if (text === '') {
+                yield* this._yieldDeepSeekStreamChunk(chunk, runManager);
+                continue;
+            }
+            tokensBuffer += text;
+            while (tokensBuffer !== '') {
+                if (isThinking) {
+                    const thinkEndIndex = tokensBuffer.indexOf(thinkEndTag);
+                    if (thinkEndIndex !== -1) {
+                        const thoughtContent = tokensBuffer.substring(0, thinkEndIndex);
+                        if (thoughtContent !== '') {
+                            yield* this._yieldDeepSeekReasoningText(chunk, thoughtContent, runManager);
+                        }
+                        tokensBuffer = tokensBuffer.substring(thinkEndIndex + thinkEndTag.length);
+                        isThinking = false;
+                        continue;
+                    }
+                    const splitIndex = this._getDeepSeekPartialTagSplitIndex(tokensBuffer, thinkEndTag);
+                    if (splitIndex !== -1) {
+                        const safeToYield = tokensBuffer.substring(0, splitIndex);
+                        if (safeToYield !== '') {
+                            yield* this._yieldDeepSeekReasoningText(chunk, safeToYield, runManager);
+                        }
+                        tokensBuffer = tokensBuffer.substring(splitIndex);
+                        break;
+                    }
+                    yield* this._yieldDeepSeekReasoningText(chunk, tokensBuffer, runManager);
+                    tokensBuffer = '';
+                    break;
+                }
+                const thinkStartIndex = tokensBuffer.indexOf(thinkStartTag);
+                if (thinkStartIndex !== -1) {
+                    const beforeThink = tokensBuffer.substring(0, thinkStartIndex);
+                    if (beforeThink !== '') {
+                        yield* this._yieldDeepSeekStreamChunk(this._createDeepSeekStreamChunk(chunk, beforeThink), runManager);
+                    }
+                    tokensBuffer = tokensBuffer.substring(thinkStartIndex + thinkStartTag.length);
+                    isThinking = true;
+                    continue;
+                }
+                const splitIndex = this._getDeepSeekPartialTagSplitIndex(tokensBuffer, thinkStartTag);
+                if (splitIndex !== -1) {
+                    const safeToYield = tokensBuffer.substring(0, splitIndex);
+                    if (safeToYield !== '') {
+                        yield* this._yieldDeepSeekStreamChunk(this._createDeepSeekStreamChunk(chunk, safeToYield), runManager);
+                    }
+                    tokensBuffer = tokensBuffer.substring(splitIndex);
+                    break;
+                }
+                yield* this._yieldDeepSeekStreamChunk(this._createDeepSeekStreamChunk(chunk, tokensBuffer), runManager);
+                tokensBuffer = '';
+                break;
+            }
+        }
+        if (tokensBuffer === '') {
+            return;
+        }
+        if (isThinking) {
+            yield* this._yieldDeepSeekStreamChunk(new outputs.ChatGenerationChunk({
+                message: new messages.AIMessageChunk({
+                    content: '',
+                    additional_kwargs: {
+                        reasoning_content: tokensBuffer,
+                    },
+                }),
+                text: '',
+            }), runManager);
+            return;
+        }
+        yield* this._yieldDeepSeekStreamChunk(new outputs.ChatGenerationChunk({
+            message: new messages.AIMessageChunk({
+                content: tokensBuffer,
+            }),
+            text: tokensBuffer,
+        }), runManager);
+    }
+    async *_streamResponseChunksFromReasoningMessages(messages$1, options) {
+        const params = {
+            ...this.invocationParams(options, { streaming: true }),
+            stream: true,
+        };
+        const messagesMapped = this._convertDeepSeekMessages(messages$1);
+        const streamIterable = await this.completionWithRetry({
+            ...params,
+            messages: messagesMapped,
+        }, {
+            signal: options.signal,
+            ...options.options,
+        });
+        let defaultRole;
+        let usage;
+        for await (const data of streamIterable) {
+            if (options.signal?.aborted === true) {
+                throw new Error('AbortError');
+            }
+            if (data.usage != null) {
+                usage = data.usage;
+            }
+            if (data.choices.length === 0) {
+                continue;
+            }
+            const choice = data.choices[0];
+            const { delta } = choice;
+            const messageChunk = this._convertCompletionsDeltaToBaseMessageChunk(delta, data, defaultRole);
+            defaultRole = delta.role ?? defaultRole;
+            if (typeof messageChunk.content !== 'string') {
+                continue;
+            }
+            const messageText = messageChunk.content;
+            const newTokenIndices = {
+                prompt: options.promptIndex ?? 0,
+                completion: choice.index,
+            };
+            const generationInfo = { ...newTokenIndices };
+            if (choice.finish_reason != null) {
+                Object.assign(generationInfo, {
+                    finish_reason: choice.finish_reason,
+                    system_fingerprint: data.system_fingerprint,
+                    model_name: data.model,
+                    service_tier: data.service_tier,
+                });
+            }
+            if (this.logprobs === true) {
+                Object.assign(generationInfo, { logprobs: choice.logprobs });
+            }
+            const generationChunk = new outputs.ChatGenerationChunk({
+                message: messageChunk,
+                text: messageText,
+                generationInfo,
+            });
+            yield generationChunk;
+        }
+        if (usage != null) {
+            const usageMetadata = createUsageMetadata(usage);
+            const generationChunk = new outputs.ChatGenerationChunk({
+                message: new messages.AIMessageChunk({
+                    content: '',
+                    response_metadata: {
+                        usage: { ...usage },
+                    },
+                    usage_metadata: usageMetadata,
+                }),
+                text: '',
+                generationInfo: {
+                    prompt: 0,
+                    completion: 0,
+                },
+            });
+            yield generationChunk;
+        }
+        if (options.signal?.aborted === true) {
+            throw new Error('AbortError');
+        }
+    }
+    _createDeepSeekStreamChunk(chunk, content, additionalKwargs, text = content) {
+        if (!(chunk.message instanceof messages.AIMessageChunk)) {
+            return new outputs.ChatGenerationChunk({
+                message: new messages.AIMessageChunk({
+                    content,
+                    additional_kwargs: additionalKwargs ?? chunk.message.additional_kwargs,
+                    response_metadata: chunk.message.response_metadata,
+                    id: chunk.message.id,
+                }),
+                text,
+                generationInfo: chunk.generationInfo,
+            });
+        }
+        const message = chunk.message;
+        return new outputs.ChatGenerationChunk({
+            message: new messages.AIMessageChunk({
+                content,
+                additional_kwargs: additionalKwargs ?? message.additional_kwargs,
+                response_metadata: message.response_metadata,
+                tool_calls: message.tool_calls,
+                tool_call_chunks: message.tool_call_chunks,
+                id: message.id,
+            }),
+            text,
+            generationInfo: chunk.generationInfo,
+        });
+    }
+    _createDeepSeekReasoningStreamChunk(chunk, reasoningContent) {
+        return this._createDeepSeekStreamChunk(chunk, '', {
+            ...chunk.message.additional_kwargs,
+            reasoning_content: reasoningContent,
+        }, '');
+    }
+    async *_yieldDeepSeekReasoningText(chunk, reasoningContent, runManager) {
+        yield* this._yieldDeepSeekStreamChunk(this._createDeepSeekReasoningStreamChunk(chunk, reasoningContent), runManager);
+    }
+    async *_yieldDeepSeekStreamChunk(chunk, runManager) {
+        yield chunk;
+        await runManager?.handleLLMNewToken(chunk.text, this._getDeepSeekTokenIndices(chunk), undefined, undefined, undefined, { chunk });
+    }
+    _getDeepSeekTokenIndices(chunk) {
+        const prompt = chunk.generationInfo?.prompt;
+        const completion = chunk.generationInfo?.completion;
+        if (typeof prompt === 'number' && typeof completion === 'number') {
+            return { prompt, completion };
+        }
+        return undefined;
+    }
+    _getDeepSeekPartialTagSplitIndex(text, tag) {
+        for (let i = tag.length - 1; i >= 1; i--) {
+            if (text.endsWith(tag.substring(0, i))) {
+                return text.length - i;
+            }
+        }
+        return -1;
     }
 }
 class ChatMoonshot extends ChatOpenAI {