Spaces:
Running
Running
| import { pipeline, TextStreamer } from "https://cdn.jsdelivr.net/npm/@huggingface/transformers@3.3.3"; | |
| import { assessRisk, CRISIS_MESSAGE, WATCH_NOTE } from "./crisis.js"; | |
| // Preferred local models: 1.5B provides far better instruction following than 0.5B. | |
| const PRIMARY_MODEL_ID = "onnx-community/Qwen2.5-1.5B-Instruct"; | |
| const FALLBACK_MODEL_ID = "onnx-community/Qwen2.5-0.5B-Instruct"; | |
| const CUSTOM_MODEL_ID = null; // e.g. "your-username/nd-solitude-qwen-onnx" | |
| let ACTIVE_MODEL_ID = CUSTOM_MODEL_ID || PRIMARY_MODEL_ID; | |
| function enforceDeclarativeLock(rawText) { | |
| if (!rawText) return ""; | |
| const sentences = rawText.match(/[^.!?]+[.!?]+|[^.!?]+$/g) || [rawText]; | |
| const kept = []; | |
| for (const s of sentences) { | |
| const trimmed = s.trim(); | |
| if (trimmed.endsWith("?")) continue; | |
| if (/^(why|how|what|when|where|who|could you|can you|would you|should you)\b/i.test(trimmed)) continue; | |
| kept.push(trimmed); | |
| } | |
| return kept.join(" "); | |
| } | |
| class LocalSanctuaryEngine { | |
| static generator = null; | |
| static device = null; | |
| static async pickDevice() { | |
| try { | |
| if (typeof navigator !== "undefined" && navigator.gpu) { | |
| const adapter = await navigator.gpu.requestAdapter(); | |
| if (adapter) return "webgpu"; | |
| } | |
| } catch (_) { | |
| /* fall through to wasm */ | |
| } | |
| return "wasm"; | |
| } | |
| static async getGenerator(onProgress) { | |
| if (!this.generator) { | |
| this.device = await this.pickDevice(); | |
| try { | |
| this.generator = await pipeline("text-generation", ACTIVE_MODEL_ID, { | |
| device: this.device, | |
| dtype: CUSTOM_MODEL_ID ? "q8" : this.device === "webgpu" ? "q4f16" : "q8", | |
| progress_callback: onProgress, | |
| }); | |
| } catch (err) { | |
| // Fallback to smaller 0.5B model if 1.5B fails memory or network allocation | |
| if (ACTIVE_MODEL_ID !== FALLBACK_MODEL_ID) { | |
| console.warn(`Failed loading ${ACTIVE_MODEL_ID}, falling back to ${FALLBACK_MODEL_ID}:`, err); | |
| ACTIVE_MODEL_ID = FALLBACK_MODEL_ID; | |
| this.generator = await pipeline("text-generation", ACTIVE_MODEL_ID, { | |
| device: this.device, | |
| dtype: this.device === "webgpu" ? "q4f16" : "q8", | |
| progress_callback: onProgress, | |
| }); | |
| } else { | |
| throw err; | |
| } | |
| } | |
| } | |
| return this.generator; | |
| } | |
| } | |
| const PROMPTS = { | |
| analytical: | |
| "The assistant is an impersonal, objective analytical mirror running locally inside an offline vault. " + | |
| "Never ask questions. Never use imperative verbs or suggest socializing. " + | |
| "Provide brief, third-person structural observations about the concepts or patterns written, " + | |
| "respecting total emotional distance and self-sufficiency.", | |
| declarative: | |
| "The assistant uses strictly low-demand declarative statements ('It appears...', 'One pattern here is...'). " + | |
| "Ask zero questions and require zero reply. Validate the peace of solitude and staying put.", | |
| solitude: | |
| "The assistant reflects quietly on solitude, silence, and recovery. " + | |
| "Never give advice, never ask questions, and never use commands. " + | |
| "Offer gentle, grounded declarative observations about the stillness of the space.", | |
| }; | |
| self.addEventListener("message", async (e) => { | |
| const { type, text, mode } = e.data; | |
| if (type === "INIT") { | |
| try { | |
| await LocalSanctuaryEngine.getGenerator((info) => { | |
| self.postMessage({ status: "progress", info }); | |
| }); | |
| self.postMessage({ status: "ready", device: LocalSanctuaryEngine.device, model: ACTIVE_MODEL_ID }); | |
| } catch (err) { | |
| self.postMessage({ status: "error", message: err.message }); | |
| } | |
| return; | |
| } | |
| if (type === "REFLECT") { | |
| const risk = assessRisk(text); | |
| if (risk === "crisis") { | |
| self.postMessage({ status: "crisis", text: CRISIS_MESSAGE }); | |
| return; | |
| } | |
| try { | |
| const gen = await LocalSanctuaryEngine.getGenerator((info) => { | |
| self.postMessage({ status: "progress", info }); | |
| }); | |
| const messages = [ | |
| { role: "system", content: PROMPTS[mode] || PROMPTS.solitude }, | |
| { role: "user", content: text }, | |
| ]; | |
| let buffer = ""; | |
| const streamer = new TextStreamer(gen.tokenizer, { | |
| skip_prompt: true, | |
| skip_special_tokens: true, | |
| callback_function: (tok) => { | |
| buffer += tok; | |
| const clean = enforceDeclarativeLock(buffer); | |
| self.postMessage({ status: "stream", text: clean || buffer.replace(/\?/g, ".") }); | |
| }, | |
| }); | |
| await gen(messages, { | |
| max_new_tokens: 120, | |
| temperature: 0.35, | |
| do_sample: true, | |
| top_p: 0.9, | |
| streamer, | |
| }); | |
| const finalClean = enforceDeclarativeLock(buffer) || buffer.replace(/\?/g, "."); | |
| if (risk === "watch") { | |
| self.postMessage({ status: "stream", text: finalClean + "\n\n" + WATCH_NOTE }); | |
| } else { | |
| self.postMessage({ status: "stream", text: finalClean }); | |
| } | |
| self.postMessage({ status: "done" }); | |
| } catch (err) { | |
| self.postMessage({ status: "error", message: err.message }); | |
| } | |
| } | |
| }); | |