clementine-agent 1.0.70 → 1.0.71

package/dist/brain/adapters/pdf.d.ts

@@ -1,8 +1,14 @@
 /**
- * Clementine — PDF adapter (text layer only).
+ * Clementine — PDF adapter.
  *
- * Yields one RawRecord per PDF page. pdf-parse concatenates pages with
- * a form-feed separator (\f), so we split on that after extraction.
+ * Fast path: pdf-parse extracts the text layer. Yields one RawRecord per
+ * page (pdf-parse concatenates pages with \f).
+ *
+ * OCR fallback: for image-only / scanned PDFs, pdf-parse returns empty
+ * text. We then ask Claude Code to read the PDF itself — its built-in
+ * Read tool handles PDFs natively (including vision for scanned pages),
+ * and the call goes through the Agent SDK so it works with the user's
+ * OAuth session (no separate ANTHROPIC_API_KEY required).
  */
 import type { RawRecord } from '../../types.js';
 export declare function parsePdf(filePath: string): AsyncIterable<RawRecord>;

package/dist/brain/adapters/pdf.js

@@ -1,13 +1,20 @@
 /**
- * Clementine — PDF adapter (text layer only).
+ * Clementine — PDF adapter.
  *
- * Yields one RawRecord per PDF page. pdf-parse concatenates pages with
- * a form-feed separator (\f), so we split on that after extraction.
+ * Fast path: pdf-parse extracts the text layer. Yields one RawRecord per
+ * page (pdf-parse concatenates pages with \f).
+ *
+ * OCR fallback: for image-only / scanned PDFs, pdf-parse returns empty
+ * text. We then ask Claude Code to read the PDF itself — its built-in
+ * Read tool handles PDFs natively (including vision for scanned pages),
+ * and the call goes through the Agent SDK so it works with the user's
+ * OAuth session (no separate ANTHROPIC_API_KEY required).
  */
 import { readFileSync } from 'node:fs';
 import path from 'node:path';
 import pdfParse from 'pdf-parse';
 import { contentHash } from './common.js';
+import { MODELS } from '../../config.js';
 export async function* parsePdf(filePath) {
     let buf;
     try {
@@ -29,7 +36,30 @@ export async function* parsePdf(filePath) {
     const pages = splitPages(result.text);
     const hasAnyText = pages.some((p) => p.trim().length > 0);
     if (!hasAnyText) {
-        throw new Error(`PDF ${path.basename(filePath)} has no extractable text — likely image-only (OCR is not supported). Re-export with a text layer or transcribe it first.`);
+        // Image-only / scanned PDF — fall back to Claude's native PDF reading.
+        const ocrPages = await ocrPdfViaClaude(filePath);
+        if (ocrPages.length === 0) {
+            throw new Error(`PDF ${path.basename(filePath)} has no extractable text and OCR returned nothing. The file may be corrupt or empty.`);
+        }
+        for (let i = 0; i < ocrPages.length; i++) {
+            const pageText = ocrPages[i].trim();
+            if (!pageText)
+                continue;
+            yield {
+                externalId: `pdf-ocr-${hint}-p${i + 1}-${contentHash(pageText)}`,
+                content: pageText,
+                rawPayload: pageText,
+                metadata: {
+                    adapter: 'pdf',
+                    extraction: 'claude-ocr',
+                    source_file: filePath,
+                    page: i + 1,
+                    total_pages: ocrPages.length,
+                    content_hash: contentHash(pageText),
+                },
+            };
+        }
+        return;
     }
     for (let i = 0; i < pages.length; i++) {
         const pageText = pages[i].trim();
@@ -41,6 +71,7 @@ export async function* parsePdf(filePath) {
             rawPayload: pageText,
             metadata: {
                 adapter: 'pdf',
+                extraction: 'text-layer',
                 source_file: filePath,
                 page: i + 1,
                 total_pages: result.numpages,
@@ -50,6 +81,50 @@ export async function* parsePdf(filePath) {
         };
     }
 }
+/**
+ * OCR fallback via the Claude Agent SDK. Asks Claude Code to Read the PDF
+ * and transcribe every page verbatim, separated by \f. Returns one string
+ * per page. Empty array on failure (caller decides how to handle).
+ */
+async function ocrPdfViaClaude(filePath) {
+    try {
+        const { query } = await import('@anthropic-ai/claude-agent-sdk');
+        const stream = query({
+            prompt: `Read the PDF at ${JSON.stringify(filePath)} using the Read tool. Transcribe every page's text verbatim — preserve the reading order, headings, lists, and paragraphs exactly as they appear. Separate pages with the form-feed character (\\f). Do NOT summarize, paraphrase, add commentary, or wrap in code fences. Output only the transcribed text.`,
+            options: {
+                model: MODELS.haiku,
+                maxTurns: 4, // Read tool call + response (a few turns of thinking is fine)
+                systemPrompt: 'You are a faithful OCR transcriber. Copy text exactly as written. When the PDF has images or scans, read the text from them using vision. Never invent content.',
+                // Claude Code's built-in Read tool handles PDFs (text + vision)
+                allowedTools: ['Read'],
+                permissionMode: 'bypassPermissions',
+                settingSources: [],
+            },
+        });
+        let text = '';
+        for await (const message of stream) {
+            if (message.type === 'assistant') {
+                const content = message
+                    .message?.content ?? [];
+                for (const block of content) {
+                    if (block.type === 'text' && typeof block.text === 'string') {
+                        text += block.text;
+                    }
+                }
+            }
+            else if (message.type === 'result') {
+                break;
+            }
+        }
+        const cleaned = text.trim();
+        if (cleaned.length < 20)
+            return [];
+        return splitPages(cleaned);
+    }
+    catch {
+        return [];
+    }
+}
 /** pdf-parse inserts \f between pages. Fall back to paragraph-size chunks if not. */
 function splitPages(text) {
     if (text.includes('\f'))

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "clementine-agent",
-  "version": "1.0.70",
+  "version": "1.0.71",
   "description": "Clementine — Personal AI Assistant (TypeScript)",
   "type": "module",
   "main": "dist/index.js",