Base de Conocimiento Personal — RAG sobre tus Notas, Diario y PDFs

What's in this pack

This is the stack you build when you want an AI to read your notes — daily journal, meeting notes, half-finished essays, the PDFs you've been hoarding since grad school — and answer questions about them without uploading any of it to a vendor's cloud.

It's a different problem from chatting with a frontier model. There the goal is reasoning power; here the goal is recall over your private corpus. You don't need GPT-5 — you need plumbing: a notes app that stores plain Markdown, an indexer that can read every format you've ever saved, an MCP or agent bridge that lets the AI query the vault, and a chat layer to ask questions.

This pack is not the same as our local-first-ai pack, which assembles a general private AI rig (Ollama, Open WebUI, ComfyUI, image gen, transcription). This one is narrowly scoped to the personal RAG layer: where your notes live, how they get indexed, how an agent reaches them.

Install in this order

1. Logseq — outliner-style, privacy-first notes app. Local Markdown files, optional encrypted sync, daily journal as a first-class concept. Start here if you don't already have a notes app you trust. The reason to pick Logseq over Notion or Bear: every note is a plain file on your disk that any tool downstream can read.
2. Obsidian Agent Client — if you're already an Obsidian user, this plugin brings ACP-based agent integrations (Claude Code, Codex, Gemini) directly into the editor. Your vault becomes the context window for the agent. Install via BRAT or releases.
3. Obsidian MCP Tools — the other Obsidian path: instead of bringing agents into Obsidian, expose Obsidian to Claude Desktop as a local MCP server. Claude can now query your vault, search notes, run dataview queries. Pick this if Claude Desktop is your daily driver.
4. joplin-mcp — same idea for Joplin users. MCP server that gives any Claude-Desktop-compatible client read/write access to your Joplin notes. Joplin's optional E2EE plus local-only MCP means the notes never leave the device.
5. Cherry Studio Knowledge Base — the all-in-one option. A desktop AI client with a local RAG engine that ingests 50+ formats (PDF, Markdown, DOCX, EPUB, web pages, even Notion exports), embeds them locally, and serves them to whichever LLM you wire up. If you want one app, not five, start here.
6. Trilium Notes — hierarchical KB with scripting, attribute search, and a true tree structure. Picks up where Logseq's flat-graph model leaves off. The right call if you think in outlines and want every note to be queryable as structured data.
7. Blinko — self-hosted personal AI note-taking with RAG already baked in. You don't wire up an indexer; the app comes with one. Closer to "a private NotebookLM" than "Obsidian with plugins". Trade-off: less customizable, more turn-key.
8. Zotero — research source manager. The bridge between your notes and the academic / PDF world. Auto-extracts metadata from papers, builds a searchable library, exposes a local API that downstream RAG tools (Cherry Studio, custom scripts) can index.
9. Paperless-ngx — self-hosted document management with OCR. The piece most knowledge-base setups forget: every paper bank statement, contract, and receipt scanned, OCR'd, tagged, indexed. Now your AI can answer "what was my electric bill in March 2024?" without you opening a single PDF.
10. Memos — lightweight self-hosted note-taking. Twitter-style short captures, tagged and searchable. The journaling-by-microblog layer that catches the thoughts too small to belong in Logseq but too important to lose.

How they fit together

  ┌────────────────────────────────────────────────────────┐
  │  Your private corpus (everything stays on disk)        │
  └────────────────────────────────────────────────────────┘
     │           │              │              │
  Logseq      Obsidian        Joplin         Trilium
  Memos       (or Blinko, all-in-one)
     │           │              │              │
     │           ├─ Obsidian Agent Client (ACP in editor)
     │           ├─ Obsidian MCP Tools  ──► Claude Desktop
     │                          │
     │                          └─ joplin-mcp  ──► Claude Desktop
     │
     └──► Cherry Studio Knowledge Base  ──┐
                                          │
  Zotero (papers) ─────────────────────►  ├──► Local LLM / Claude / GPT
  Paperless-ngx (scanned PDFs)    ─────►  │     (your choice of model)
                                          ▼
                                  Chat with your KB

The pattern: a notes app holds the raw text, a bridge layer (MCP server, agent plugin, or built-in RAG engine) makes it queryable, and an LLM client does the asking. You don't need every tool in this pack — pick the row matching the notes app you already use, then add Zotero and Paperless-ngx if your knowledge isn't all in Markdown.

Tradeoffs you'll hit

• MCP bridge vs in-app agent vs all-in-one — Three architectures, all valid. MCP bridge (Obsidian MCP Tools, joplin-mcp) keeps your existing notes app and lets Claude Desktop reach in; best for power users with one favorite client. In-app agent (Obsidian Agent Client, Blinko) puts the AI inside the editor; best when you want the answer next to the source. All-in-one (Cherry Studio, Blinko) bundles indexer + chat + model in one app; best when you don't want to maintain three tools.
• Embedding quality vs setup pain — Cheap path: use the notes app's built-in indexer (Blinko, Cherry Studio). Better-quality path: BGE-M3 or nomic-embed-text via Ollama, then point a custom RAG pipeline at your vault. Most people overestimate how much retrieval quality matters for personal notes — the corpus is small and you usually remember roughly where the answer lives. Start simple.
• Frontier model vs local model for the chat layer — RAG with Claude 4.5 or GPT-5 over your notes gives the best synthesis quality. RAG with a local Llama 3.1 8B over your notes keeps the journal 100% private but synthesis is noticeably weaker. Hybrid is fine: route the retrieval locally (so the embeddings of your notes never leave the machine), then send only the top-3 chunks plus your question to a frontier model.
• PDFs are the silent killer — Markdown notes index in seconds. Scanned PDFs need OCR (Paperless-ngx), academic PDFs need layout-aware extraction (Zotero handles citation metadata; for full-text RAG you may want GROBID or unstructured.io on top). Plan a separate pass for any non-text source.

Common pitfalls

• Indexing everything at once — A 10,000-file vault embedded against a poor chunking strategy gives you 10,000 useless 4-line snippets. Start with one sub-folder (say, your meeting notes from this year), measure retrieval quality, then expand.
• Daily-journal noise drowns out actual knowledge — If you also dump 2 KB of daily standup notes into the same vault, retrieval will surface yesterday's todos every time. Separate the corpora: daily journal in Logseq, evergreen notes in Obsidian, and only index the evergreen vault.
• MCP servers asking for too much — MCP gives the AI a lot of power. Read each server's permissions before installing. joplin-mcp and Obsidian MCP Tools both default to read+write; start with read-only until you trust the workflow.
• OCR quality on bad scans — Paperless-ngx is good but not magic. Phone-camera receipts at 30 degrees come back as gibberish. Use a flatbed scanner or the iOS Notes scan feature (auto-flattens) for anything you'll actually want to retrieve.
• Forgetting to back up the vault — The whole point is that this is your second brain. Encrypt and back it up to your own storage (Syncthing, a private rsync target, Joplin's server). Don't trust a single laptop with a decade of journal entries.