create-entity-app-server 0.0.3

package/src/app/plugins/llm/providers/ollama.ts

@@ -0,0 +1,237 @@
+/**
+ * LLM 클라이언트 — Ollama 로컬 LLM 드라이버
+ *
+ * Ollama /api/chat, /api/embeddings API 호출
+ */
+
+import type {
+    LlmClient,
+    LlmProviderConfig,
+    ChatRequest,
+    ChatResponse,
+    StreamChunk,
+    EmbedRequest,
+    EmbedResponse,
+} from "../types/index.ts";
+
+export class OllamaClient implements LlmClient {
+    private readonly cfg: LlmProviderConfig;
+    private readonly baseURL: string;
+    private readonly timeout: number;
+
+    /** OllamaClient 인스턴스를 초기화한다 */
+    constructor(cfg: LlmProviderConfig) {
+        this.cfg = cfg;
+        this.baseURL = (cfg.base_url || "http://localhost:11434").replace(
+            /\/+$/,
+            "",
+        );
+        this.timeout = (cfg.timeout_sec || 120) * 1000;
+    }
+
+    /** 프로바이더 이름을 반환한다 */
+    name(): string {
+        return "ollama";
+    }
+
+    /** 클라이언트 리소스를 정리한다 */
+    close(): void {
+        /* no-op */
+    }
+
+    // ─── Chat ────────────────────────────────────────────────────────────
+
+    /** 채팅 완성 API를 호출한다 */
+    async chat(req: ChatRequest, signal?: AbortSignal): Promise<ChatResponse> {
+        const body = this.buildBody(req, false);
+
+        const resp = await fetch(`${this.baseURL}/api/chat`, {
+            method: "POST",
+            headers: { "Content-Type": "application/json" },
+            body: JSON.stringify(body),
+            signal: signal ?? AbortSignal.timeout(this.timeout),
+        });
+
+        const text = await resp.text();
+        if (!resp.ok) {
+            throw new Error(`Ollama: status ${resp.status}: ${text}`);
+        }
+
+        const result: OllamaChatResponse = JSON.parse(text);
+        const chatResp: ChatResponse = {
+            content: result.message?.content ?? "",
+            finishReason: result.done_reason ?? "",
+            model: result.model ?? this.cfg.model,
+        };
+
+        if (result.eval_count > 0 || result.prompt_eval_count > 0) {
+            chatResp.usage = {
+                promptTokens: result.prompt_eval_count ?? 0,
+                completionTokens: result.eval_count ?? 0,
+                totalTokens:
+                    (result.prompt_eval_count ?? 0) + (result.eval_count ?? 0),
+            };
+        }
+
+        return chatResp;
+    }
+
+    // ─── Stream ──────────────────────────────────────────────────────────
+
+    /** 스트리밍 채팅 API를 호출한다 */
+    async *chatStream(
+        req: ChatRequest,
+        signal?: AbortSignal,
+    ): AsyncGenerator<StreamChunk> {
+        const body = this.buildBody(req, true);
+
+        const resp = await fetch(`${this.baseURL}/api/chat`, {
+            method: "POST",
+            headers: { "Content-Type": "application/json" },
+            body: JSON.stringify(body),
+            signal: signal ?? AbortSignal.timeout(this.timeout),
+        });
+
+        if (!resp.ok) {
+            const text = await resp.text();
+            throw new Error(`Ollama stream: status ${resp.status}: ${text}`);
+        }
+
+        const reader = resp.body!.getReader();
+        const decoder = new TextDecoder();
+        let buffer = "";
+
+        try {
+            while (true) {
+                const { done, value } = await reader.read();
+                if (done) break;
+
+                buffer += decoder.decode(value, { stream: true });
+                const lines = buffer.split("\n");
+                buffer = lines.pop()!;
+
+                for (const line of lines) {
+                    if (!line.trim()) continue;
+
+                    try {
+                        const event: OllamaChatResponse = JSON.parse(line);
+
+                        const chunk: StreamChunk = {
+                            content: event.message?.content ?? "",
+                            done: event.done ?? false,
+                        };
+
+                        if (event.done) {
+                            chunk.finishReason = event.done_reason ?? "stop";
+                            if (
+                                event.eval_count > 0 ||
+                                event.prompt_eval_count > 0
+                            ) {
+                                chunk.usage = {
+                                    promptTokens: event.prompt_eval_count ?? 0,
+                                    completionTokens: event.eval_count ?? 0,
+                                    totalTokens:
+                                        (event.prompt_eval_count ?? 0) +
+                                        (event.eval_count ?? 0),
+                                };
+                            }
+                        }
+
+                        yield chunk;
+                        if (event.done) return;
+                    } catch {
+                        // JSON parse 실패 — 무시
+                    }
+                }
+            }
+        } finally {
+            reader.releaseLock();
+        }
+    }
+
+    // ─── Embed ───────────────────────────────────────────────────────────
+
+    /** 텍스트 임베딩 API를 호출한다 */
+    async embed(
+        req: EmbedRequest,
+        signal?: AbortSignal,
+    ): Promise<EmbedResponse> {
+        const model = req.model || this.cfg.embed_model || this.cfg.model;
+
+        // Ollama는 단일 텍스트 임베딩만 지원
+        const input = req.input.length > 0 ? req.input[0] : "";
+        const body = { model, prompt: input };
+
+        const resp = await fetch(`${this.baseURL}/api/embeddings`, {
+            method: "POST",
+            headers: { "Content-Type": "application/json" },
+            body: JSON.stringify(body),
+            signal: signal ?? AbortSignal.timeout(this.timeout),
+        });
+
+        const text = await resp.text();
+        if (!resp.ok) {
+            throw new Error(`Ollama embed: status ${resp.status}: ${text}`);
+        }
+
+        const result = JSON.parse(text);
+
+        return {
+            embeddings: result.embedding ? [result.embedding] : [],
+            model,
+            totalTokens: 0,
+        };
+    }
+
+    // ─── Internal ────────────────────────────────────────────────────────
+
+    /** 요청 바디를 구성한다 */
+    private buildBody(
+        req: ChatRequest,
+        stream: boolean,
+    ): Record<string, unknown> {
+        const model = this.cfg.model || "llama3";
+        const messages: Array<{ role: string; content: string }> = [];
+
+        if (req.system) {
+            messages.push({ role: "system", content: req.system });
+        }
+        for (const m of req.messages) {
+            messages.push({ role: m.role, content: m.content });
+        }
+
+        const body: Record<string, unknown> = {
+            model,
+            messages,
+            stream,
+        };
+
+        const options: Record<string, unknown> = {};
+        const temp = req.temperature ?? this.cfg.temperature;
+        if (temp != null && temp > 0) options.temperature = temp;
+
+        const maxTokens = req.maxTokens ?? this.cfg.max_tokens;
+        if (maxTokens != null && maxTokens > 0) options.num_predict = maxTokens;
+
+        if (Object.keys(options).length > 0) {
+            body.options = options;
+        }
+
+        if (req.jsonMode || this.cfg.json_mode) {
+            body.format = "json";
+        }
+
+        return body;
+    }
+}
+
+// ─── Response types ──────────────────────────────────────────────────────────
+
+interface OllamaChatResponse {
+    model: string;
+    done: boolean;
+    done_reason?: string;
+    message?: { role: string; content: string };
+    prompt_eval_count: number;
+    eval_count: number;
+}

package/src/app/plugins/llm/providers/openai.ts

@@ -0,0 +1,244 @@
+/**
+ * LLM 클라이언트 — OpenAI 드라이버
+ *
+ * OpenAI Chat Completions / Embeddings API 호출.
+ * vllm, localai, lmstudio, llamacpp 등 OpenAI-호환 서버도 이 드라이버로 동작합니다.
+ */
+
+import type {
+    LlmClient,
+    LlmProviderConfig,
+    ChatRequest,
+    ChatResponse,
+    StreamChunk,
+    TokenUsage,
+    EmbedRequest,
+    EmbedResponse,
+} from "../types/index.ts";
+
+export class OpenAIClient implements LlmClient {
+    private readonly cfg: LlmProviderConfig;
+    private readonly baseURL: string;
+    private readonly timeout: number;
+
+    /** OpenAIClient 인스턴스를 초기화한다 */
+    constructor(cfg: LlmProviderConfig) {
+        this.cfg = cfg;
+        this.baseURL = (cfg.base_url || "https://api.openai.com").replace(
+            /\/+$/,
+            "",
+        );
+        this.timeout = (cfg.timeout_sec || 30) * 1000;
+    }
+
+    /** 프로바이더 이름을 반환한다 */
+    name(): string {
+        return this.cfg.driver || "openai";
+    }
+
+    /** 클라이언트 리소스를 정리한다 */
+    close(): void {
+        /* no-op */
+    }
+
+    // ─── Chat ────────────────────────────────────────────────────────────
+
+    /** 채팅 완성 API를 호출한다 */
+    async chat(req: ChatRequest, signal?: AbortSignal): Promise<ChatResponse> {
+        const body = this.buildBody(req, false);
+        const resp = await this.doRequest("/v1/chat/completions", body, signal);
+
+        const choice = resp.choices?.[0];
+        const chatResp: ChatResponse = {
+            content: choice?.message?.content ?? "",
+            finishReason: choice?.finish_reason ?? "",
+            model: resp.model ?? this.cfg.model,
+        };
+
+        if (resp.usage) {
+            chatResp.usage = {
+                promptTokens: resp.usage.prompt_tokens ?? 0,
+                completionTokens: resp.usage.completion_tokens ?? 0,
+                totalTokens: resp.usage.total_tokens ?? 0,
+            };
+        }
+
+        return chatResp;
+    }
+
+    // ─── Stream ──────────────────────────────────────────────────────────
+
+    /** 스트리밍 채팅 API를 호출한다 */
+    async *chatStream(
+        req: ChatRequest,
+        signal?: AbortSignal,
+    ): AsyncGenerator<StreamChunk> {
+        const body = this.buildBody(req, true);
+        const response = await fetch(`${this.baseURL}/v1/chat/completions`, {
+            method: "POST",
+            headers: this.buildHeaders(),
+            body: JSON.stringify(body),
+            signal: signal ?? AbortSignal.timeout(this.timeout),
+        });
+
+        if (!response.ok) {
+            const text = await response.text();
+            throw new Error(
+                `OpenAI stream: status ${response.status}: ${text}`,
+            );
+        }
+
+        const reader = response.body!.getReader();
+        const decoder = new TextDecoder();
+        let buffer = "";
+
+        try {
+            while (true) {
+                const { done, value } = await reader.read();
+                if (done) break;
+
+                buffer += decoder.decode(value, { stream: true });
+                const lines = buffer.split("\n");
+                buffer = lines.pop()!;
+
+                for (const line of lines) {
+                    const trimmed = line.trim();
+                    if (!trimmed.startsWith("data: ")) continue;
+                    const data = trimmed.slice(6);
+                    if (data === "[DONE]") {
+                        yield { content: "", done: true };
+                        return;
+                    }
+
+                    try {
+                        const event = JSON.parse(data);
+                        const delta = event.choices?.[0]?.delta;
+                        const finishReason = event.choices?.[0]?.finish_reason;
+                        const isDone =
+                            finishReason != null && finishReason !== "";
+
+                        const chunk: StreamChunk = {
+                            content: delta?.content ?? "",
+                            done: isDone,
+                        };
+
+                        if (isDone) {
+                            chunk.finishReason = finishReason;
+                        }
+                        if (event.usage) {
+                            chunk.usage = {
+                                promptTokens: event.usage.prompt_tokens ?? 0,
+                                completionTokens:
+                                    event.usage.completion_tokens ?? 0,
+                                totalTokens: event.usage.total_tokens ?? 0,
+                            };
+                        }
+
+                        yield chunk;
+                        if (isDone) return;
+                    } catch {
+                        // JSON parse 실패 — 무시
+                    }
+                }
+            }
+        } finally {
+            reader.releaseLock();
+        }
+    }
+
+    // ─── Embed ───────────────────────────────────────────────────────────
+
+    /** 텍스트 임베딩 API를 호출한다 */
+    async embed(
+        req: EmbedRequest,
+        signal?: AbortSignal,
+    ): Promise<EmbedResponse> {
+        const model = req.model || this.cfg.embed_model || this.cfg.model;
+        const body = { input: req.input, model };
+
+        const resp = await this.doRequest("/v1/embeddings", body, signal);
+
+        const embeddings: number[][] = (resp.data ?? []).map(
+            (d: { embedding: number[] }) => d.embedding,
+        );
+
+        return {
+            embeddings,
+            model: resp.model ?? model,
+            totalTokens: resp.usage?.total_tokens ?? 0,
+        };
+    }
+
+    // ─── Internal ────────────────────────────────────────────────────────
+
+    /** 요청 헤더를 구성한다 */
+    private buildHeaders(): Record<string, string> {
+        const h: Record<string, string> = {
+            "Content-Type": "application/json",
+        };
+        if (this.cfg.api_key) {
+            h["Authorization"] = `Bearer ${this.cfg.api_key}`;
+        }
+        return h;
+    }
+
+    /** 요청 바디를 구성한다 */
+    private buildBody(
+        req: ChatRequest,
+        stream: boolean,
+    ): Record<string, unknown> {
+        const messages: Array<{ role: string; content: string }> = [];
+
+        if (req.system) {
+            messages.push({ role: "system", content: req.system });
+        }
+        for (const m of req.messages) {
+            messages.push({ role: m.role, content: m.content });
+        }
+
+        const body: Record<string, unknown> = {
+            model: this.cfg.model,
+            messages,
+            stream,
+        };
+
+        const temp = req.temperature ?? this.cfg.temperature;
+        if (temp != null && temp > 0) body.temperature = temp;
+
+        const maxTokens = req.maxTokens ?? this.cfg.max_tokens;
+        if (maxTokens != null && maxTokens > 0) body.max_tokens = maxTokens;
+
+        const topP = req.topP ?? this.cfg.top_p;
+        if (topP != null && topP > 0) body.top_p = topP;
+
+        if (req.jsonMode || this.cfg.json_mode) {
+            body.response_format = { type: "json_object" };
+        }
+
+        if (req.stop?.length) body.stop = req.stop;
+        if (stream) body.stream_options = { include_usage: true };
+
+        return body;
+    }
+
+    /** API 요청을 실행하고 응답을 파싱한다 */
+    private async doRequest(
+        path: string,
+        body: unknown,
+        signal?: AbortSignal,
+    ): Promise<any> {
+        const resp = await fetch(`${this.baseURL}${path}`, {
+            method: "POST",
+            headers: this.buildHeaders(),
+            body: JSON.stringify(body),
+            signal: signal ?? AbortSignal.timeout(this.timeout),
+        });
+
+        const text = await resp.text();
+        if (!resp.ok) {
+            throw new Error(`OpenAI: status ${resp.status}: ${text}`);
+        }
+
+        return JSON.parse(text);
+    }
+}

package/src/app/plugins/llm/routes.ts

@@ -0,0 +1,73 @@
+/**
+ * LLM 라우트 테이블
+ * 등록 경로: /api/v1/llm/*
+ */
+
+import type { FastifyInstance } from "fastify";
+import * as h from "./handlers";
+/** LLM 라우트를 등록한다 */ export default async function llmRoutes(
+    app: FastifyInstance,
+): Promise<void> {
+    // 서비스 존재 확인
+    app.addHook("preHandler", h.ensureService);
+
+    // ── 채팅 ──
+    app.post("/chat", h.chat);
+    app.post("/chat/stream", h.chatStream);
+
+    // ── 대화 세션 ──
+    app.post("/conversations", h.createConversation);
+    app.post("/conversations/:seq/messages", h.sendMessage);
+    app.get("/conversations", h.listConversations);
+    app.get("/conversations/:seq", h.getConversation);
+    app.patch("/conversations/:seq", h.updateConversation);
+    app.delete("/conversations/:seq", h.deleteConversation);
+
+    // ── RAG ──
+    app.post("/rag/documents", h.ragUploadDocument);
+    app.get("/rag/documents", h.ragListDocuments);
+    app.delete("/rag/documents/:id", h.ragDeleteDocument);
+    app.post("/rag/search", h.ragSearch);
+    app.post("/rag/chat", h.ragChat);
+    app.post("/rag/chat/stream", h.ragChatStream);
+    app.post("/rag/rebuild-index", h.ragRebuildIndex);
+
+    // ── 프로바이더 / 사용량 ──
+    app.get("/providers", h.listProviders);
+    app.get("/usage", h.getUsage);
+    app.get("/usage/summary", h.getUsageSummary);
+
+    // ── 캐시 ──
+    app.get("/cache/stats", h.getCacheStats);
+    app.delete("/cache", h.clearCache);
+
+    // ── 프롬프트 템플릿 ──
+    // 정적 라우트가 모두 등록된 후 마지막에 배치.
+    // Fastify는 정적 세그먼트를 파라미터보다 항상 우선 매칭하므로
+    // /chat, /rag/chat 등 기존 경로와 충돌하지 않는다.
+    app.get("/templates", h.listPromptTemplates);
+
+    // ── 챗봇 관리 ──
+    app.get("/chatbots", h.listChatbots);
+    app.post("/chatbots", h.createChatbot);
+    app.get("/chatbots/:seq", h.getChatbot);
+    app.patch("/chatbots/:seq", h.updateChatbot);
+    app.delete("/chatbots/:seq", h.deleteChatbot);
+
+    // ── 챗봇 채팅 ──
+    app.post("/chatbots/:seq/chat", h.chatbotChat);
+    app.post("/chatbots/:seq/chat/stream", h.chatbotChatStream);
+
+    // ── 챗봇 세션 ──
+    app.get("/chatbots/:seq/sessions", h.listChatbotSessions);
+    app.delete("/chatbots/:seq/sessions/:sessionSeq", h.deleteChatbotSession);
+
+    // ── Profile Memory ──
+    app.get("/profiles", h.listProfiles);
+    app.post("/profiles", h.upsertProfile);
+    app.delete("/profiles/:seq", h.deleteProfile);
+
+    // ── 템플릿 폴백 (:name) ── 반드시 정적 라우트 모두 등록 후에 배치
+    app.post("/:name/chat", h.templateChat);
+    app.post("/:name/chat/stream", h.templateChatStream);
+}