obsidian_ollama/README.md

# Obsidian Ollama Plugin

A plugin that integrates [Ollama](https://ollama.ai) with Obsidian, allowing you to chat with local AI models, search your vault context, and use AI tools like creating files.

## Features

- **Chat with Ollama models** directly in Obsidian with streaming responses
- **Vault context search** — the assistant can reference your notes via semantic (RAG) or keyword search
- **Agent Modes** — selectable chat modes (Ask, Edit, Organize, Research, Workflow) that change available tools, system prompts, and preview behaviour
- **Tool integration** — create, read, search, append, edit, rename, move, delete notes, and insert wiki-links
- **Structured Memory** — persist conversation summaries, user preferences, and learned facts across sessions
- **Tool Telemetry** — track which tools were called, which notes were searched, and LLM token usage
- **Semantic/RAG vault indexing** — automatically index your vault into a vector database for intelligent retrieval
- **Semantic response cache** — repeated or similar queries are answered instantly without hitting the model
- **Workflow Engine** — execute multi-step AI workflows via `/workflow` commands
- **Auto-Organizer** — AI-powered auto-tagging and auto-linking with dry-run preview and folder scoping
- **Obsidian MetadataCache integration** — frontmatter, tags, links, and headings are read via Obsidian's built-in cache instead of raw regex parsing
- Customisable model, URL, cache, and memory settings

## Prerequisites

1. **Install Ollama**: Follow the instructions at [ollama.ai](https://ollama.ai)
2. **Start Ollama**: `ollama serve`
3. **Pull a chat model**: `ollama pull llama3` (or any other model you prefer)

### Optional — Vault Semantic Index (RAG)

The vault semantic index automatically indexes your Obsidian notes into a local [ChromaDB](https://www.trychroma.com) vector database. When you ask a question, the plugin performs semantic search against your notes and includes the most relevant passages as context for the AI.

1. **Install ChromaDB**:
   ```bash
   pip install chromadb
   ```
2. **Start ChromaDB**:
   ```bash
   chroma run --host localhost --port 8000
   ```
3. **Pull an embedding model**:
   ```bash
   ollama pull nomic-embed-text
   ```
4. Enable the vault semantic index in the plugin settings and configure the ChromaDB URL.

### Optional — Semantic Cache

The semantic cache stores responses in a local [ChromaDB](https://www.trychroma.com) vector database. When you ask a question that is semantically similar to one already cached, the stored answer is returned immediately instead of calling the model.

1. **Install ChromaDB**:
   ```bash
   pip install chromadb
   ```
2. **Start ChromaDB**:
   ```bash
   chroma run --host localhost --port 8000
   ```
3. **Pull an embedding model** (used to generate vectors for cache lookups):
   ```bash
   ollama pull nomic-embed-text
   ```
4. Enable the cache in the plugin settings and configure the ChromaDB URL.

## Installation

### Quick install (recommended)

Use the included install script. It handles dependency installation, building, and copying the plugin into your vault:

```bash
# Clone or download this repository, then run:
./install.sh /path/to/your/obsidian/vault
```

The script will:
- Install npm dependencies (excluding Ollama — you install that separately)
- Compile the TypeScript plugin
- Copy the built plugin into `<vault>/.obsidian/plugins/ollama-plugin/`

### Manual install

If you prefer to install manually:

```bash
npm install
npm run build
```

Then copy the plugin into your vault:

```bash
mkdir -p /path/to/vault/.obsidian/plugins/ollama-plugin
cp manifest.json /path/to/vault/.obsidian/plugins/ollama-plugin/
cp main.js /path/to/vault/.obsidian/plugins/ollama-plugin/
cp styles.css /path/to/vault/.obsidian/plugins/ollama-plugin/
# Remove old dist/ from previous installs (no longer needed with bundling)
rm -rf /path/to/vault/.obsidian/plugins/ollama-plugin/dist
```

> **Note:** The plugin is now bundled into a single `main.js` via esbuild. The `obsidian` npm package is a dev-only type stub — Obsidian provides its own API at runtime. The `chromadb` client library is also bundled into `main.js`, so no extra `node_modules` copy is needed for the semantic cache feature.

### After installation

1. Restart Obsidian (or reload: `Ctrl+Shift+P` → "Reload app without saving")
2. Go to **Settings → Community plugins** → enable **Ollama Plugin**
3. Configure the plugin at **Settings → Ollama Settings**

## Configuration

Open **Settings → Ollama Settings** to configure the plugin.

| Setting | Default | Description |
|---------|---------|-------------|
| Ollama URL | `http://localhost:11434` | Base URL of your Ollama instance |
| Chat Model | `deepseek-v4-flash` | Model used for normal chat, Ask mode, and Research mode |
| Agent Model | `glm-5.1` | Model used for Edit, Organize, Workflow, and auto-organizer tasks |
| **Default Agent Mode** | `Ask` | Default chat mode (Ask, Edit, Organize, Research, Workflow) |
| Vault Search Limit | `5` | Maximum number of vault entries to include in context |
| Max Context Length | `8000` | Maximum characters of vault content sent to the AI per message |
| Max Message History | `50` | Maximum number of messages kept in conversation history |
| **Enable Vault Semantic Index** | Off | Index vault notes into a vector DB for semantic/RAG search |
| Vault Index ChromaDB URL | `http://localhost:8000` | URL of your ChromaDB instance for the vault index |
| Vault Index Embedding Model | `nomic-embed-text` | Ollama model used to generate vault embeddings |
| Vault Index Similarity Threshold | `0.75` | Minimum cosine similarity (0–1) for a vault search hit |
| Rebuild Vault Index | — | Button to rebuild the entire vault semantic index |
| Clear Vault Index | — | Button to delete all indexed vault notes |
| Enable Semantic Cache | Off | Cache responses for fast repeated queries |
| ChromaDB URL | `http://localhost:8000` | URL of your running ChromaDB instance |
| Cache Embedding Model | `nomic-embed-text` | Ollama model used to generate cache embeddings |
| Cache Similarity Threshold | `0.85` | Minimum cosine similarity (0–1) for a cache hit |
| Clear Semantic Cache | — | Button to wipe all cached responses |
| **Enable Auto-Tagging** | Off | Automatically suggest and apply tags to untagged notes |
| Max Tags Per Note | `5` | Maximum tags to generate per note |
| Normalize Tags | On | Normalize generated tags against existing vault vocabulary |
| Target Folder (Auto-Tag) | — | Restrict auto-tagging to a specific folder |
| **Enable Auto-Linking** | Off | Add "Related Notes" sections based on semantic similarity |
| Max Links Per Note | `3` | Maximum related note links to insert |
| Target Folder (Auto-Link) | — | Restrict auto-linking to a specific folder |
| Dry Run Mode (Auto-Link) | Off | Preview proposed links without applying them |
| **Enable Structured Memory** | On | Inject remembered context from past sessions into prompts |
| Max Conversation Summaries | `10` | Maximum past conversation summaries to retain |
| Max User Preferences | `20` | Maximum user preferences to retain |
| Max Learned Facts | `50` | Maximum learned facts to retain |
| Clear Structured Memory | — | Button to delete all stored memory |
| **Enable Tool Telemetry** | On | Record tool calls, searches, and LLM token counts |
| Max Telemetry Entries | `100` | Maximum telemetry events to retain |
| Clear Tool Telemetry | — | Button to delete all recorded telemetry |

## Usage

1. Open the chat view via the command palette (`Ctrl+P` → "Open Ollama Chat") or the ribbon icon
2. Type your message in the input box
3. Press **Enter** or click **Send** to send your message
4. Press **Shift+Enter** to insert a line break
5. Click **New Chat** to start a fresh conversation

### Agent Modes

The chat view includes a mode selector dropdown. Each mode changes the assistant's behaviour:

| Mode | Tools Available | Preview Required | Use Case |
|------|----------------|------------------|----------|
| **Ask** | Read, Search | No | Answer questions using vault context |
| **Edit** | All tools | Yes | Create, modify, and manage notes |
| **Organize** | Read, Search, Frontmatter, Rename, Move, Link | Yes | Tag, rename, move, and link notes |
| **Research** | Read, Search | No | Deep vault search and synthesis |
| **Workflow** | None (uses `/workflow`) | No | Execute multi-step AI workflows |

When a mode requires preview (Edit, Organize), write operations like `create_note` or `delete_note` show a card with a before/after diff and **Apply** / **Cancel** buttons. Ask and Research modes execute write tools immediately without preview.

### Workflows

Type `/workflow` followed by a description to trigger the workflow engine. The AI will generate a multi-step workflow plan, then execute it step-by-step. Example:

```
/workflow Find all notes tagged "meeting", summarise them, and create a "Meeting Summary" note
```

### Auto-Organizer

Use the command palette to trigger:
- **Auto-Tag Untagged Notes** — AI generates tags for notes missing tags
- **Auto-Link Related Notes** — AI inserts "Related Notes" sections with wiki-links

Both features support:
- **Dry-run mode** — preview proposed changes without modifying the vault
- **Target folder** — restrict processing to a specific folder and its subfolders
- **Tag normalisation** — match generated tags against existing vault vocabulary

### Commands

| Command | Description |
|---------|-------------|
| Open Ollama Chat | Open the chat sidebar |
| Clear Semantic Cache | Delete all cached responses |
| Clear Vault Index | Delete all indexed vault notes |
| Rebuild Vault Index | Rebuild the vault semantic index from scratch |
| Auto-Tag Untagged Notes | Run the auto-tagger |
| Auto-Link Related Notes | Run the auto-linker |
| Clear Structured Memory | Delete all conversation summaries, preferences, and facts |
| Clear Tool Telemetry | Delete all recorded telemetry events |

## Semantic Cache Behaviour

- The cache is **bypassed** when tool calls are involved (e.g. file creation), since those requests have side effects.
- Responses are stored against the last user message in the conversation. If a new query is sufficiently similar (above the configured threshold), the cached response is returned.
- Re-asking the same question updates the existing cache entry rather than creating a duplicate.
- Use the **Clear Semantic Cache** button in settings to remove all stored responses (for example after switching embedding models).

## Vault Context

When you send a message, the plugin searches your vault for relevant notes and includes them as context. If the **Vault Semantic Index** is enabled, search is performed via semantic/RAG retrieval using vector embeddings. Otherwise, it falls back to a weighted keyword search:

- **Headings** — 5x weight
- **Frontmatter title** — 3x weight
- **Frontmatter tags** — 2.5x weight
- **First paragraph** — 1.5x weight
- **General content** — 1x weight

The plugin also pulls in:
- **Explicit mentions** — notes referenced via `[[...]]` wikilinks in the message
- **Open note** — the currently active note
- **Selected text** — text selected in the active editor
- **Backlinks / Outlinks** — notes that link to / from the open note
- **Related notes** — semantically similar notes (requires vault semantic index)

The plugin automatically watches your vault for changes (create, modify, delete, rename) and updates the semantic index in real time when enabled.

Frontmatter, tags, links, and headings are resolved using Obsidian's built-in `metadataCache` API for accuracy and performance.

## Tools

The assistant has access to a suite of tools that interact with your vault. Available tools depend on the current **Agent Mode**:

| Tool | Description | Mode |
|------|-------------|------|
| `create_note` / `create_file` | Create a new markdown file | Edit |
| `read_vault_file` | Read the contents of a note | All |
| `search_vault_files` | Keyword-search vault files by path | All |
| `append_to_note` | Append text to the end of a note | Edit |
| `replace_note_section` | Replace content under a specific heading | Edit |
| `update_frontmatter` | Add, update, or remove frontmatter fields | Edit, Organize |
| `rename_note` | Rename a note file | Edit, Organize |
| `move_note` | Move a note to a different folder | Edit, Organize |
| `delete_note` | Delete a note | Edit |
| `insert_link` | Insert a `[[wiki-link]]` into a note | Edit, Organize |

Paths are validated for safety: no `.obsidian`/`.git` access, no path traversal (`..`), no absolute paths, and a 200-character limit.

## Structured Memory

When **Enable Structured Memory** is on, the plugin remembers context across sessions by storing three kinds of data in Obsidian's plugin data JSON:

- **Conversation Summaries** — After each assistant reply, a brief summary (topic + key points) is saved
- **User Preferences** — Statements like "I prefer dark mode" or "My favourite colour is blue" are extracted and stored
- **Learned Facts** — Simple facts mentioned in conversation (e.g., "Obsidian is a note-taking app") and vault folder paths are remembered

These are injected as a system message at the start of every LLM call, so the assistant "remembers" context from previous sessions. Limits and clear controls are available in settings.

## Tool Telemetry

When **Enable Tool Telemetry** is on, the plugin records:

- **Tool calls** — which tool, arguments, success/failure, result summary, and duration
- **LLM calls** — model, estimated prompt/completion/total tokens, and duration
- **Vault searches** — query, number of results, and matched note paths

Telemetry is stored locally in Obsidian's plugin data. The settings tab shows a **Recent Activity** summary of the last 10 events. Use **Clear Tool Telemetry** to wipe the history.

> **Note:** Token counts are exact when Ollama provides `prompt_eval_count` and `eval_count` in its response; otherwise they are estimated from character count (÷4 approximation).

## Supported Models

Any Ollama-supported model works. Popular choices:

- `llama3`
- `llama2`
- `mistral`
- `codellama`
- and many more — see [ollama.com/library](https://ollama.com/library)

## Development

```bash
npm install
npm run build
npm test          # 520+ unit tests across 21 test suites
npm run lint      # ESLint check
```

The project uses TypeScript, Jest, and esbuild. Obsidian APIs are mocked in `__mocks__/obsidian.ts` for testing.

## Troubleshooting

| Symptom | Likely cause | Fix |
|---------|-------------|-----|
| Plugin doesn't appear in Obsidian | Install script was not run or failed | Run `./install.sh /path/to/vault` and reload Obsidian |
| Cannot connect to Ollama | Ollama is not running | Run `ollama serve` |
| Model not found | Model not pulled | Run `ollama pull <model>` |
| "Invalid response format" error | Ollama returned a non-JSON response (e.g., proxy error page) | Check that Ollama is healthy at the configured URL |
| Semantic cache unavailable (notice shown) | ChromaDB is not running, or the ChromaDB URL is wrong | Start ChromaDB (`chroma run`) and verify the URL in settings |
| Cache always misses | Similarity threshold is too high, or the embedding model was changed | Lower the threshold or click **Clear Semantic Cache** and let the cache rebuild |
| Slow first response after enabling cache | Embedding model not yet pulled | Run `ollama pull nomic-embed-text` (or the model you configured) |
| Permission issues | Vault write permissions | Check that your Obsidian vault has proper write permissions |
| Structured memory not showing up | Memory was just cleared or is empty | Have a few conversations — summaries are generated after each assistant reply |
| Tool telemetry not recording | Telemetry is disabled or max entries is 0 | Enable **Tool Telemetry** in settings and set **Max Telemetry Entries** > 0 |

## Security

- File paths are validated to prevent access to `.obsidian/` and `.git/` directories
- Path traversal attempts (`..`) are blocked
- Absolute paths and Windows drive letters are rejected
- Maximum path length is enforced (200 characters)

## License

MIT License

Copyright (c) 2024 Flo Egger

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.