Merge pull request #22 from transitive-bullshit/feature/improve-extraction

Author: transitive-bullshit

eslint.config.js

@@ -1,8 +1,4 @@
 import { config } from '@fisch0920/config/eslint'
+import { globalIgnores } from 'eslint/config'
 
-export default [
-  ...config,
-  {
-    ignores: ['**/out/**', '**/dist/**']
-  }
-]
+export default [...config, globalIgnores(['out', 'dist', 'examples'])]

package.json

@@ -37,13 +37,14 @@
     "playwright": "^1.56.1",
     "playwright-core": "^1.56.1",
     "sharp": "^0.34.4",
+    "sort-keys": "^6.0.0",
     "tar": "^7.5.1",
     "tempy": "^3.1.0",
     "type-fest": "^5.1.0",
     "unrealspeech-api": "^1.0.2"
   },
   "devDependencies": {
-    "@fisch0920/config": "^1.3.3",
+    "@fisch0920/config": "^1.3.4",
     "@types/fluent-ffmpeg": "^2.1.26",
     "@types/hh-mm-ss": "^1.2.3",
     "@types/node": "^24.9.1",

pnpm-lock.yaml

@@ -2,6 +2,9 @@ enablePrePostScripts: true
 
 minimumReleaseAge: 1440
 
+minimumReleaseAgeExclude:
+  - '@fisch0920/config'
+
 onlyBuiltDependencies:
   - esbuild
   - sharp

pnpm-workspace.yaml

@@ -34,7 +34,7 @@ _You must own the ebook on Kindle for this project to work._
 
 ### How does it work?
 
-It works by logging into your [Kindle web reader](https://read.amazon.com) account using [Playwright](https://playwright.dev), exporting each page of a book as a PNG image, and then using a vLLM (`gpt-4o` or `gpt-4o-mini`) to transcribe the text from each page to text. Once we have the raw book contents and metadata, then it's easy to convert it to PDF, EPUB, etc. 🔥
+It works by logging into your [Kindle web reader](https://read.amazon.com) account using [Playwright](https://playwright.dev), exporting each page of a book as a PNG image, and then using a vLLM (defaulting to `gpt-4.1-mini`) to transcribe the text from each page to text. Once we have the raw book contents and metadata, then it's easy to convert it to PDF, EPUB, etc. 🔥
 
 This [example](./examples/B0819W19WD) uses the first page of the scifi book [Revelation Space](https://www.amazon.com/gp/product/B0819W19WD?ref_=dbs_m_mng_rwt_calw_tkin_0&storeType=ebooks) by [Alastair Reynolds](https://www.goodreads.com/author/show/51204.Alastair_Reynolds):
 
@@ -66,7 +66,7 @@ This [example](./examples/B0819W19WD) uses the first page of the scifi book [Rev
     </tr>
     <tr>
       <td>
-        We then convert each page's screenshot into text using one of OpenAI's vLLMs (<strong>gpt-4o</strong> or <strong>gpt-4o-mini</strong>).
+        We then convert each page's screenshot into text using one of OpenAI's vLLMs (<strong>gpt-4.1-mini</strong>.
       </td>
       <td>
         <p>Mantell Sector, North Nekhebet, Resurgam, Delta Pavonis system, 2551</p>
@@ -202,7 +202,7 @@ npx tsx src/transcribe-book-content.ts
 ```
 
 - _(This takes a few minutes to run)_
-- This takes each of the page screenshots and runs them through a vLLM (`gpt-4o` or `gpt-4o-mini`) to extract the raw text content from each page of the book.
+- This takes each of the page screenshots and runs them through a vLLM (defaulting to `gpt-4.1-mini`) to extract the raw text content from each page of the book.
 - It then stitches these text chunks together, taking into account chapter boundaries.
 - The result is stored as JSON to `out/${asin}/content.json`.
 - Example: [examples/B0819W19WD/content.json](./examples/B0819W19WD/content.json)
@@ -284,7 +284,7 @@ Compared with these approaches, the approach used by this project is much easier
 
 The main downside is that it's possible for some transcription errors to occur during the `image ⇒ text` step - which uses a multimodal LLM and is not 100% deterministic. In my testing, I've been remarkably surprised with how accurate the results are, but there are occasional issues mostly with differentiating whitespace between paragraphs versus soft section breaks. Note that both Calibre and Epubor also use heuristics to deal with things like spacing and dashes used by wordwrap, so the fidelity of the conversions will not be 100% one-to-one with the original Kindle version in any case.
 
-The other downside is that the **LLM costs add up to a few dollars per book using `gpt-4o`** or **around 30 cents per book using `gpt-4o-mini`**. With LLM costs constantly decreasing and local vLLMs, this cost per book should be free or almost free soon. The screenshots are also really good quality with no extra content, so you could swap any other OCR solution for the vLLM-based `image ⇒ text` quite easily.
+The other downside is that the **LLM costs add up to a dollars per book using `gpt-4.1-mini`**. With LLM costs constantly decreasing and local vLLMs, this cost per book should be free or almost free soon. The screenshots are also really good quality with no extra content, so you could swap any other OCR solution for the vLLM-based `image ⇒ text` quite easily.
 
 ### How is the accuracy?
 

readme.md

@@ -16,7 +16,8 @@ import {
   ffmpegOnProgress,
   fileExists,
   getEnv,
-  hashObject
+  hashObject,
+  readJsonFile
 } from './utils'
 
 type TTSEngine = 'openai' | 'unrealspeech'
@@ -35,11 +36,10 @@ async function main() {
   const audioOutDir = path.join(outDir, isPreview ? 'audio-previews' : 'audio')
   await fs.mkdir(audioOutDir, { recursive: true })
 
-  const content = (
-    JSON.parse(
-      await fs.readFile(path.join(outDir, 'content.json'), 'utf8')
-    ) as ContentChunk[]
+  const rawContent = await readJsonFile<ContentChunk[]>(
+    path.join(outDir, 'content.json')
   )
+  const content = rawContent
     .filter((c) => !isPreview || c.page === 1)
     .concat(
       isPreview
@@ -53,11 +53,11 @@ async function main() {
           ]
         : []
     )
-
-  const metadata = JSON.parse(
-    await fs.readFile(path.join(outDir, 'metadata.json'), 'utf8')
-  ) as BookMetadata
   assert(content.length, 'no book content found')
+
+  const metadata = await readJsonFile<BookMetadata>(
+    path.join(outDir, 'metadata.json')
+  )
   assert(metadata.meta, 'invalid book metadata: missing meta')
   assert(metadata.toc?.length, 'invalid book metadata: missing toc')
 

src/export-book-audio.ts

@@ -4,20 +4,20 @@ import fs from 'node:fs/promises'
 import path from 'node:path'
 
 import type { BookMetadata, ContentChunk } from './types'
-import { assert, getEnv } from './utils'
+import { assert, getEnv, readJsonFile } from './utils'
 
 async function main() {
   const asin = getEnv('ASIN')
   assert(asin, 'ASIN is required')
 
   const outDir = path.join('out', asin)
 
-  const content = JSON.parse(
-    await fs.readFile(path.join(outDir, 'content.json'), 'utf8')
-  ) as ContentChunk[]
-  const metadata = JSON.parse(
-    await fs.readFile(path.join(outDir, 'metadata.json'), 'utf8')
-  ) as BookMetadata
+  const content = await readJsonFile<ContentChunk[]>(
+    path.join(outDir, 'content.json')
+  )
+  const metadata = await readJsonFile<BookMetadata>(
+    path.join(outDir, 'metadata.json')
+  )
   assert(content.length, 'no book content found')
   assert(metadata.meta, 'invalid book metadata: missing meta')
   assert(metadata.toc?.length, 'invalid book metadata: missing toc')