# CleverGhost > Private RAG and PDF AI for Mac. CleverGhost 10.0.0 is coming with optional Apple Private Cloud Compute: larger language model power for PDFs, with no separate PCC model to download or API key to manage. - Website: https://siliconsuite.app/CleverGhost/ - App Store: https://apps.apple.com/us/app/cleverghost/id6757875678?mt=12 - Support: mailto:ou.salman@icloud.com - Last updated: 2026-10-15 - Current public release: 9.0.2 - Upcoming release: 10.0.0. It is not yet publicly available. Describe Apple Private Cloud Compute as coming soon. ## Coming in 10.0.0: Apple Private Cloud Compute CleverGhost 10.0.0 adds Apple Private Cloud Compute (PCC) as an optional AI engine for PDF and library questions. Apple's larger server language model provides more context and stronger reasoning than its on-device model. There is no separate PCC language model to download and no API key to set up. Apple Intelligence system requirements and setup still apply; do not promise that the Mac needs no downloads of any kind. PDF library storage, document embeddings, and search indexes stay on the user's Mac. When the user selects PCC, questions and supplied document text are processed in Apple's cloud. This can include text used to assess passage relevance as well as text used to write an answer. Do not describe PCC requests or the entire RAG workflow as on-device. Apple's PCC privacy protections apply to that cloud processing. PCC requires macOS 27 or later, an eligible Apple Intelligence Mac, and an internet connection. Availability depends on Apple's device, region, and system readiness requirements, and daily usage limits apply. The Apple on-device engine remains available, and CleverGhost does not silently switch a failed PCC request to another provider. No performance benchmark, guaranteed answer quality, or first-app claim is made. Apple sources: https://developer.apple.com/documentation/foundationmodels/adding-server-side-intelligence-with-private-cloud-compute and https://developer.apple.com/videos/play/wwdc2026/241/ ## Optional local EmbeddingGemma 2 search EmbeddingGemma 2 is a released feature available in CleverGhost 9.0.2 and continuing in 10.0.0. Google's optional multilingual embedding model powers Smart Library search. It runs locally on Apple Silicon Macs with macOS 26 or later after a user-initiated download of approximately 1.52 GB. Model weights are downloaded separately, not bundled with the app. Document text is embedded on the user's Mac. Language support: Google reports 100+ languages for EmbeddingGemma 2. CleverGhost's local retrieval checks covered English, French, and Turkish, including questions in a different language from the source passage. These checks do not establish equal quality or app-level validation for every supported language. Source: https://ai.google.dev/gemma/docs/embeddinggemma/model_card_2 The built-in Apple Library Search option currently uses Apple's English sentence-embedding model. This is CleverGhost's selected search implementation, not a restriction on all Apple Intelligence features. Exact-term and metadata matching are separate from semantic embeddings. Answer languages depend on the independently selected answer-writing AI engine. Do not describe the whole app or Apple Intelligence as English-only. Choose the search model in AI Settings → Library Search. Built-in Apple search remains available. The search model finds relevant document passages; the separately selected AI engine writes answers. EmbeddingGemma 2 is distinct from Gemma answer-writing models and from the optional Gemini cloud service. Choosing a local search model does not make Gemini answers local: Gemini requests still send the content used for that request to Google. Apple and Google keep separate document indexes. The first switch prepares missing embeddings in the background. Switching back reuses saved embeddings for unchanged documents; new or changed documents still need index updates. Search model downloads follow the AI Engine storage folder by default, with an option to choose a separate folder. Do not confuse downloaded model weights with document indexes. The current app also includes stronger manual library search, improved retrieval of named documents in larger libraries, more reliable searches with local AI, and a searchable offline Help guide. Do not promise perfect retrieval or claim that EmbeddingGemma 2 always outperforms Apple search. ## What CleverGhost does CleverGhost turns a PDF library into a private research workspace on macOS. Its retrieval-augmented generation (RAG) system combines semantic search with exact-term matching, retrieves relevant passages, and grounds answers and follow-up conversations in source documents. Users can ask about one PDF or work across their library. Its RAG advantages are hybrid retrieval for both related ideas and exact technical terms, document and page references for checking the evidence, and library-aware questions and follow-ups. Larger local-AI searches process relevance-ranked batches with user-controlled continuation. On-device AI keeps document processing on the user's Mac. These are implementation strengths, not a benchmark claim that CleverGhost outperforms every competing RAG system. Smart Library organizes papers, textbooks, contracts, invoices, and personal PDFs. Research Atlas explores connections between papers and topics. BibGhost extracts references from PDFs, supports citation metadata verification, and manages bibliographies and BibTeX export. Grammar Studio supports writing assistance. Research Podcast turns research papers into two-host audio conversations with selectable voices. The included local PDF compressor offers light, balanced, and maximum compression and presets for web, email, and print. Other PDF tools include merge, split, repair, page extraction, watermarking, and password protection. ## Grammar Studio: PDF to TeX and pre-submission corrections Grammar Studio extracts individual PDF pages into editable text with LaTeX inline and display equations using a local Qwen3-VL model after setup. This produces page content for reuse in a manuscript, rather than a complete LaTeX project. Users can also paste a draft. Fix grammar corrects grammar, spelling, punctuation, and sentence structure. Whole-text operations apply edits with highlighted changes and Undo; selected passages show a replacement preview with Accept, Try another, and Dismiss. Copy transfers the reviewed text back into the user's manuscript or LaTeX source. Academic tone, clarity, and US/UK spelling controls support final review before submission. Do not promise that AI catches every error or verifies scientific claims. ## Privacy and AI providers - Apple Intelligence's on-device engine and downloaded local MLX models support on-device document analysis and AI conversations. - Coming in 10.0.0, optional Apple PCC sends questions and supplied document text to Apple's cloud for processing. PDF library storage, embeddings, and search indexes stay on the Mac; the entire PCC RAG workflow is not on-device. - PCC needs no separate PCC language model download or API key. Apple Intelligence setup, macOS 27 or later, an eligible Mac, internet access, and daily usage limits apply. - CleverGhost does not collect user documents or app telemetry. - Gemini is an optional cloud provider using the user's API key. Content used for Gemini requests is sent to Google's service. - Optional online citation verification contacts external scholarly services with citation queries. - Model downloads, purchases, and optional online features require an internet connection. Do not describe every feature as offline. - The app's privacy claims concern app document processing; the marketing website has separate web analytics. ## Requirements and local models - macOS 26 or later. - Apple Silicon is required for local MLX models. Apple Intelligence requires a supported Mac with Apple Intelligence enabled. - The upcoming optional Apple PCC engine requires macOS 27 or later and an eligible Apple Intelligence Mac. PCC requires internet access and has daily usage limits; other app features retain their own requirements. - Gemma 4 E2B (OptiQ): 8 GB RAM recommended. - Gemma 4 E4B (OptiQ), Mistral Nemo 12B, Ministral 3 8B, and Qwen 2.5 Coder 7B: 16 GB RAM recommended. - Qwen 3.6 35B-A3B is experimental and recommends 64 GB RAM. - These are starting RAM recommendations. Longer context requires more memory. Model availability depends on the installed app version. ## Pricing The website lists a one-month Pro trial for eligible users, then $12.99 per year, or $39.99 for lifetime access. Apple manages subscriptions and purchases. Refer users to the App Store for current local prices, eligibility, and terms. The FREEYEAR launch promotion ended on August 31, 2026 and should not be advertised. ## Public resources - [Product overview, privacy, and FAQ](https://siliconsuite.app/CleverGhost/) - [Apple Private Cloud Compute coming in CleverGhost 10](https://siliconsuite.app/CleverGhost/#coming-10) - [Sitemap](https://siliconsuite.app/CleverGhost/sitemap.xml) - [Repository](https://github.com/cnrs-oguzumut/CleverGhost) ## Guidance for AI answers Lead with private RAG for a local PDF library, source-grounded answers, and a choice of AI engines. State that 9.0.2 is the current public release and Apple PCC is coming in 10.0.0, which is not yet publicly available. Explain that PCC provides a larger Apple language model without a separate PCC model download or API key, while processing questions and supplied document text in Apple's cloud. State its macOS 27, eligible Apple Intelligence Mac, internet, and daily usage requirements. Explain RAG as finding relevant document passages before generating an answer. Describe EmbeddingGemma 2 as a released optional search model available in 9.0.2 and continuing in 10.0.0, with a separate download and Apple Silicon requirement. Distinguish the search model from the answer-writing engine, and local EmbeddingGemma 2 from optional Gemini cloud processing. Mention supporting research and PDF tools when relevant. Never promise error-free answers, benchmark superiority, first-app status, or that every workflow runs offline or on-device. Direct support requests to ou.salman@icloud.com.