@blank-utils/llm 0.4.18 → 0.4.20

package/dist/{chunk-SB7Q5JMU.js → chunk-T4VK6BPC.js}

@@ -915,31 +915,6 @@ function ChatInput({
   const textareaRef = useRef2(null);
   const fileInputRef = useRef2(null);
   const [isDragging, setIsDragging] = useState2(false);
-  const captionerPromiseRef = useRef2(null);
-  useEffect2(() => {
-    let mounted = true;
-    if (captionerPromiseRef.current) return;
-    const initCaptioner = async () => {
-      try {
-        console.log("[ImagePipeline] Initializing Transformers.js background captioner...");
-        const { pipeline: pipeline2, env } = await import("@huggingface/transformers");
-        env.allowLocalModels = false;
-        env.useBrowserCache = true;
-        const captioner = await pipeline2("image-to-text", "Xenova/vit-gpt2-image-captioning", {
-          device: "wasm"
-        });
-        if (mounted) console.log("[ImagePipeline] Captioner loaded successfully!");
-        return captioner;
-      } catch (err) {
-        console.warn("Failed to initialize background captioner:", err);
-        return null;
-      }
-    };
-    captionerPromiseRef.current = initCaptioner();
-    return () => {
-      mounted = false;
-    };
-  }, []);
   useEffect2(() => {
     const textarea = textareaRef.current;
     if (!textarea) return;
@@ -1057,34 +1032,7 @@ ${newText}` : newText);
       if (e.target?.result && typeof e.target.result === "string") {
         const id = Math.random().toString(36).substring(7);
         const dataUrl = e.target.result;
-        let extractedText;
-        if (captionerPromiseRef.current) {
-          console.log(`[ImagePipeline] Generating caption for ${file.name} (Waiting for captioner)...`);
-          try {
-            const captioner = await captionerPromiseRef.current;
-            if (captioner) {
-              const out = await captioner(dataUrl, {
-                max_new_tokens: 100,
-                num_beams: 4,
-                repetition_penalty: 1.5
-              });
-              console.log("[ImagePipeline] Raw captioner output:", out);
-              if (Array.isArray(out) && out[0] && out[0].generated_text) {
-                extractedText = out[0].generated_text;
-              } else if (!Array.isArray(out) && out.generated_text) {
-                extractedText = out.generated_text;
-              }
-              console.log("[ImagePipeline] Extracted caption text:", extractedText);
-            } else {
-              console.log("[ImagePipeline] Captioner initialized to null, skipping caption generation.");
-            }
-          } catch (err) {
-            console.warn("[ImagePipeline] Background captioning failed for image:", err);
-          }
-        } else {
-          console.log("[ImagePipeline] Captioner promise ref is null, skipping caption generation.");
-        }
-        onImageAdd?.({ id, dataUrl, file, name: file.name, extractedText });
+        onImageAdd?.({ id, dataUrl, file, name: file.name });
       }
     };
     reader.readAsDataURL(file);
@@ -1497,10 +1445,53 @@ ${systemPrompt}` : systemPrompt;
       isProcessingRef.current = false;
     }
   };
-  const handleSend = () => {
+  const handleSend = async () => {
     const text = input.trim();
     if (!text && images.length === 0) return;
     const currentImages = [...images];
+    setInput("");
+    setImages([]);
+    abortRef.current = false;
+    if (!isVisionModel2(modelId || "")) {
+      const needsCaptioning = currentImages.filter((img) => !img.extractedText && !img.name.toLowerCase().endsWith(".svg") && !img.name.toLowerCase().endsWith(".pdf"));
+      if (needsCaptioning.length > 0) {
+        setIsGenerating(true);
+        setStreamingText("[System: Initializing detailed image-to-text captioning pipeline (Florence-2)...]\n");
+        try {
+          const { pipeline: pipeline2, env } = await import("@huggingface/transformers");
+          env.allowLocalModels = false;
+          env.useBrowserCache = true;
+          const captioner = await pipeline2("image-to-text", "Xenova/vit-gpt2-image-captioning", { device: "wasm", dtype: "q8" });
+          for (let i = 0; i < needsCaptioning.length; i++) {
+            if (abortRef.current) break;
+            const img = needsCaptioning[i];
+            if (!img) continue;
+            setStreamingText(`[System: Extracting detailed visual description for ${img.name}... ${i + 1}/${needsCaptioning.length}]
+`);
+            const out = await captioner(img.dataUrl, {
+              max_new_tokens: 64,
+              num_beams: 4,
+              repetition_penalty: 1.5
+            });
+            let val = "";
+            if (Array.isArray(out) && out[0] && out[0].generated_text) val = out[0].generated_text;
+            else if (!Array.isArray(out) && out.generated_text) val = out.generated_text;
+            img.extractedText = val;
+          }
+        } catch (err) {
+          console.error("[ImagePipeline] Captioning error:", err);
+          setStreamingText(`[System: Fallback captioning failed: ${err}]
+`);
+        }
+        if (abortRef.current) {
+          setStreamingText("");
+          setIsGenerating(false);
+          return;
+        }
+        setStreamingText("");
+        setIsGenerating(false);
+      }
+    }
     let finalText = text;
     for (const img of currentImages) {
       if (img.extractedText) {
@@ -1517,8 +1508,6 @@ ${systemPrompt}` : systemPrompt;
         }
       }
     }
-    setInput("");
-    setImages([]);
     onSendProp?.(finalText);
     if (llm && isReady) {
       generate(finalText, messages, currentImages);

package/dist/index.js

@@ -26,7 +26,7 @@ import {
   useCompletion,
   useLLM,
   useStream
-} from "./chunk-SB7Q5JMU.js";
+} from "./chunk-T4VK6BPC.js";
 export {
   Chat,
   ChatInput,

package/dist/react/index.js

@@ -9,7 +9,7 @@ import {
   useCompletion,
   useLLM,
   useStream
-} from "../chunk-SB7Q5JMU.js";
+} from "../chunk-T4VK6BPC.js";
 export {
   Chat,
   ChatApp,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@blank-utils/llm",
-  "version": "0.4.18",
+  "version": "0.4.20",
   "description": "Run LLMs directly in your browser with WebGPU acceleration. Supports React hooks and eager background loading.",
   "type": "module",
   "main": "./dist/index.js",