dsh-knowledge-library 0.2.0 → 0.3.0

package/dist/knowledge-model.js

@@ -1,6 +1,42 @@
 const unavailable = '请先在 Harness 设置中配置默认模型；独立预览可通过环境变量配置模型。';
 class ModelResponseError extends Error {
 }
+function outputLimitMessage() {
+    return '模型输出达到当前模型或服务的长度上限。插件沿用模型默认输出配置，未另设 token 上限；思考可能与正文共用额度，本次未自动重试。请检查模型配置，或分批整理后主动重试。';
+}
+async function boundedJson(response, limit = 512_000) {
+    const reader = response.body?.getReader();
+    if (!reader)
+        throw new Error('模型响应为空');
+    const chunks = [];
+    let length = 0;
+    try {
+        for (;;) {
+            const part = await reader.read();
+            if (part.done)
+                break;
+            length += part.value.length;
+            if (length > limit)
+                throw new Error('模型响应过大，请缩小整理范围');
+            chunks.push(part.value);
+        }
+    }
+    finally {
+        await reader.cancel().catch(() => { });
+    }
+    try {
+        return JSON.parse(Buffer.concat(chunks).toString('utf8'));
+    }
+    catch {
+        throw new Error('模型响应不是有效 JSON');
+    }
+}
+function checkImages(images) {
+    if (!images?.length)
+        return;
+    if (images.length > 10 || images.some(image => !image.data.byteLength || image.data.byteLength > 6 * 1024 * 1024 || !['image/png', 'image/jpeg', 'image/webp', 'image/gif'].includes(image.mediaType)) || images.reduce((sum, image) => sum + image.data.byteLength, 0) > 20 * 1024 * 1024)
+        throw new Error('图片请求超过数量、格式或字节限制');
+}
 export function harnessKnowledgeModel(ctx) {
     function services() {
         const llm = ctx.get('llm');
@@ -10,7 +46,21 @@ export function harnessKnowledgeModel(ctx) {
             throw new Error(unavailable);
         return { llm, selection };
     }
+    async function visionStatus() {
+        try {
+            const { llm, selection } = services();
+            const model = `${selection.provider} / ${selection.model}`;
+            if (!ctx.get('attachments'))
+                return { available: false, model, reason: '当前宿主未提供图片附件服务，请升级或启用附件服务' };
+            const info = await llm.resolveModelInfo(selection.provider, selection.model, AbortSignal.timeout(5000));
+            return info.inputModalities?.includes('image') ? { available: true, model } : { available: false, model, reason: '当前默认模型未声明图片输入能力，请在 Harness 中选择支持视觉的模型（如 deepseek-flash）' };
+        }
+        catch {
+            return { available: false, reason: '无法确认默认模型的图片能力，请检查 Harness 模型设置' };
+        }
+    }
     return {
+        visionStatus,
         status() {
             try {
                 const { selection } = services();
@@ -20,13 +70,31 @@ export function harnessKnowledgeModel(ctx) {
                 return { available: false, reason: unavailable };
             }
         },
-        async generate({ system, prompt, maxTokens, signal }) {
+        async generate({ system, prompt, signal, images, expectedModel }) {
+            checkImages(images);
             const { llm, selection } = services();
+            if (expectedModel !== undefined && expectedModel !== `${selection.provider} / ${selection.model}`)
+                throw new Error('模型已改变，请重新核对');
             signal.throwIfAborted();
+            const content = [{ type: 'text', text: prompt }];
+            if (images?.length) {
+                const status = await visionStatus();
+                signal.throwIfAborted();
+                if (!status.available || status.model !== `${selection.provider} / ${selection.model}`)
+                    throw new Error(status.reason || '图片模型配置已改变');
+                const attachments = ctx.get('attachments');
+                const refs = await attachments.saveImages(images);
+                signal.throwIfAborted();
+                content.push(...refs.map(attachment => ({ type: 'image', attachment })));
+            }
+            const current = services().selection;
+            if (current.provider !== selection.provider || current.model !== selection.model || current.reasoningEffort !== selection.reasoningEffort)
+                throw new Error('模型配置已改变，请重新核对');
             let text = '';
             let finished = false;
+            // Omit maxTokens so the Harness runtime materializes the model's own defaults.
             try {
-                for await (const chunk of llm.stream({ ...selection, system, messages: [{ role: 'user', content: [{ type: 'text', text: prompt }] }], maxTokens, signal })) {
+                for await (const chunk of llm.stream({ ...selection, system, messages: [{ role: 'user', content }], signal })) {
                     signal.throwIfAborted();
                     // Complete blocks are authoritative; deltas may include reasoning or retries.
                     if (chunk.type === 'block-end' && chunk.block.type === 'text')
@@ -35,7 +103,7 @@ export function harnessKnowledgeModel(ctx) {
                         throw new ModelResponseError('模型输出过长，请缩小整理范围');
                     if (chunk.type === 'finish') {
                         if (chunk.reason.kind !== 'stop')
-                            throw new ModelResponseError(chunk.reason.kind === 'max-tokens' ? '模型输出达到长度上限，请缩小整理范围后重试' : '模型调用未完成，请检查 Harness 模型设置后重试');
+                            throw new ModelResponseError(chunk.reason.kind === 'max-tokens' ? outputLimitMessage() : '模型调用未完成，请检查 Harness 模型设置后重试');
                         finished = true;
                     }
                 }
@@ -58,11 +126,45 @@ export function previewKnowledgeModel(env = process.env) {
     const key = env.KNOWLEDGE_LIBRARY_MODEL_API_KEY || env.DEEPSEEK_API_KEY;
     const model = env.KNOWLEDGE_LIBRARY_MODEL || 'deepseek-chat';
     const base = env.KNOWLEDGE_LIBRARY_MODEL_BASE_URL || 'https://api.deepseek.com';
+    let cached;
+    async function visionStatus() {
+        if (!key)
+            return { available: false, model, reason: unavailable };
+        if (cached && Date.now() - cached.time < 15_000)
+            return cached.value;
+        let value;
+        try {
+            const response = await fetch(`${base.replace(/\/$/, '')}/models`, { headers: { Authorization: `Bearer ${key}` }, redirect: 'error', signal: AbortSignal.timeout(5000) });
+            if (!response.ok) {
+                await response.body?.cancel();
+                throw new Error();
+            }
+            const data = await boundedJson(response);
+            const entry = Array.isArray(data.data) && data.data.find((item) => item.id === model);
+            value = entry?.input_modalities?.includes('image') ? { available: true, model } : { available: false, model, reason: '当前模型未声明图片输入能力，请设置支持视觉的 KNOWLEDGE_LIBRARY_MODEL（如 deepseek-flash）' };
+        }
+        catch {
+            value = { available: false, model, reason: '无法读取模型图片能力，请检查服务地址的 /models 接口和网络' };
+        }
+        cached = { time: Date.now(), value };
+        return value;
+    }
     return {
+        visionStatus,
         status: () => key ? { available: true, model } : { available: false, reason: unavailable },
-        async generate({ system, prompt, maxTokens, signal }) {
+        async generate({ system, prompt, signal, images, expectedModel }) {
+            checkImages(images);
             if (!key)
                 throw new Error(unavailable);
+            if (expectedModel !== undefined && expectedModel !== model)
+                throw new Error('模型已改变，请重新核对');
+            signal.throwIfAborted();
+            if (images?.length) {
+                const info = await visionStatus();
+                signal.throwIfAborted();
+                if (!info.available)
+                    throw new Error(info.reason);
+            }
             let response;
             try {
                 const url = new URL(`${base.replace(/\/$/, '')}/chat/completions`);
@@ -70,7 +172,7 @@ export function previewKnowledgeModel(env = process.env) {
                     throw new Error();
                 response = await fetch(url, { method: 'POST', redirect: 'error', signal,
                     headers: { 'Content-Type': 'application/json', Authorization: `Bearer ${key}` },
-                    body: JSON.stringify({ model, stream: false, max_tokens: maxTokens, messages: [{ role: 'system', content: system }, { role: 'user', content: prompt }] }) });
+                    body: JSON.stringify({ model, stream: false, messages: [{ role: 'system', content: system }, { role: 'user', content: images?.length ? [{ type: 'text', text: prompt }, ...images.map(image => ({ type: 'image_url', image_url: { url: `data:${image.mediaType};base64,${Buffer.from(image.data).toString('base64')}`, detail: 'original' } }))] : prompt }] }) });
             }
             catch {
                 signal.throwIfAborted();
@@ -80,34 +182,10 @@ export function previewKnowledgeModel(env = process.env) {
                 await response.body?.cancel();
                 throw new Error(`模型调用失败（HTTP ${response.status}），请检查模型配置或额度`);
             }
-            // Bound bytes before JSON parsing, including chunked responses.
-            const reader = response.body?.getReader();
-            if (!reader)
-                throw new Error('模型响应为空');
-            const chunks = [];
-            let length = 0;
-            try {
-                while (true) {
-                    const part = await reader.read();
-                    if (part.done)
-                        break;
-                    length += part.value.length;
-                    if (length > 512_000)
-                        throw new Error('模型响应过大，请缩小整理范围');
-                    chunks.push(part.value);
-                }
-            }
-            finally {
-                await reader.cancel().catch(() => { });
-            }
-            let data;
-            try {
-                data = JSON.parse(Buffer.concat(chunks).toString('utf8'));
-            }
-            catch {
-                throw new Error('模型响应不是有效 JSON');
-            }
+            const data = await boundedJson(response);
             const choice = data.choices?.[0];
+            if (choice?.finish_reason === 'length' || choice?.finish_reason === 'max_tokens')
+                throw new Error(outputLimitMessage());
             if (choice?.finish_reason !== 'stop' || typeof choice.message?.content !== 'string' || !choice.message.content.trim())
                 throw new Error('模型未返回完整正文，可能达到输出上限，请缩小范围后重试');
             return choice.message.content;

package/dist/public/app.css

@@ -327,6 +327,25 @@
 .kl-info .kl-proxy-option { display: flex; align-items: flex-start; gap: 8px; color: var(--kl-ink); font-size: 13px; line-height: 20px; cursor: pointer; }
 .kl-proxy-option input { flex: none; width: 16px; height: 16px; margin: 2px 0; accent-color: var(--kl-ink); }
 .kl-network-settings p { font-size: 12px; margin: 8px 0 12px; }
+.kl-image-settings { margin-top: 20px; padding-top: 16px; border-top: .5px solid var(--kl-line); }
+.kl-image-settings h3, .kl-page-images h3 { font-size: 13px; font-weight: 500; margin: 0 0 12px; }
+.kl-image-settings p, .kl-image-record p { font-size: 12px; overflow-wrap: anywhere; }
+.kl-info .kl-image-limit { display: flex; align-items: center; gap: 12px; margin: 14px 0; }
+.kl-image-limit input { width: 68px; padding: 6px 8px; border: .5px solid var(--kl-line); border-radius: 8px; background: var(--kl-bg); }
+.kl-image-policy { padding: 10px 12px; border-left: 2px solid var(--kl-link); background: var(--kl-sidebar); color: var(--kl-secondary); font-size: 12px; line-height: 1.8; margin: 0 0 16px; overflow-wrap: anywhere; }
+.kl-page-images { margin: 24px 0; padding-block: 16px; border-block: .5px solid var(--kl-line); }
+.kl-image-heading { display: flex; justify-content: space-between; align-items: center; flex-wrap: wrap; gap: 8px; }
+.kl-image-heading h3 { margin: 0; }
+.kl-image-grid { display: grid; grid-template-columns: repeat(auto-fit, minmax(min(190px, 100%), 1fr)); gap: 16px; }
+.kl-image-record { min-width: 0; display: flex; flex-direction: column; align-items: flex-start; gap: 6px; font-size: 12px; overflow-wrap: anywhere; }
+.kl-image-record > a:has(img) { width: 100%; }
+.kl-image-record img { display: block; width: 100%; height: 140px; object-fit: contain; border: .5px solid var(--kl-line); border-radius: 8px; background: var(--kl-sidebar); }
+.kl-image-record small { color: var(--kl-muted); }
+.kl-image-record a { color: var(--kl-link); }
+.kl-image-record p { margin: 0; color: var(--kl-secondary); }
+.kl-delete-dialog li { overflow-wrap: anywhere; }
+.kl-job > p { color: var(--kl-secondary); }
+.kl-job > p[role="alert"] { color: var(--kl-error); }
 .kl-error, .kl-form-error { background: var(--kl-sidebar); color: var(--kl-error); font-size: 13px; line-height: 1.7; padding: 10px 12px; border-radius: 8px; overflow-wrap: anywhere; }
 .kl-error { display: flex; align-items: center; gap: 12px; margin: 0 var(--kl-gutter) 14px; }
 .kl-error button { margin-left: auto; border: 0; background: transparent; color: inherit; display: flex; padding: 4px; }