mirror of
https://github.com/ggozad/oterm.git
synced 2026-10-09 08:43:21 +02:00
97fd157160- rag_example: rewrite for current haiku.rag. The MCP server is `haiku-rag mcp` and opens the database read-only, so documents are added with the CLI; the page described `serve`, SQLite storage and add/delete tools that no longer exist. - mcp: stdio servers accept cwd and inherit HOME, LOGNAME, PATH, SHELL, TERM and USER. - app_config: each openaiCompatible endpoint is its own provider; XDG_DATA_HOME applies everywhere but Windows; list OTERM_OLLAMA_IMAGE_MODEL and everything under OTERM_DATA_DIR. - commands: add Copy message, an Images section and the command-line options; clicking a code block copies the whole message. - installation: document the speak extra and nix updates; drop the 0.13.1 tap note. - development: uv-based setup; docs are built with Zensical. - One H1 per page; no em dashes. - README points at the CHANGELOG; drop the stale announcement banner.
77 lines
2.9 KiB
Markdown
77 lines
2.9 KiB
Markdown
# RAG with haiku.rag
|
|
|
|
<video controls>
|
|
<source src="../img/haiku-rag-demo.mp4" type="video/mp4">
|
|
</video>
|
|
|
|
[haiku.rag](https://github.com/ggozad/haiku.rag), also by `oterm`'s author, is a RAG library. It stores documents in LanceDB, searches them with hybrid (vector and full-text) search, and serves that search over MCP. Its default configuration runs the embedding model through Ollama.
|
|
|
|
## Build a database
|
|
|
|
The MCP server opens the database read-only, so documents go in through the `haiku-rag` command line:
|
|
|
|
```bash
|
|
uvx haiku-rag init --db /path/to/rag.lancedb
|
|
uvx haiku-rag add-src /path/to/report.pdf --db /path/to/rag.lancedb
|
|
uvx haiku-rag add-src https://example.com/article.html --db /path/to/rag.lancedb
|
|
uvx haiku-rag add-src /path/to/notes/ --db /path/to/rag.lancedb
|
|
```
|
|
|
|
A directory is added recursively. See the haiku.rag [command line reference](https://ggozad.github.io/haiku.rag/cli/) for everything else, including listing and deleting documents.
|
|
|
|
## Connect oterm
|
|
|
|
Add the server to `config.json` (run `oterm --data-dir` to find it):
|
|
|
|
```json
|
|
{
|
|
"mcpServers": {
|
|
"haiku-rag": {
|
|
"command": "uvx",
|
|
"args": ["haiku-rag", "mcp", "--stdio", "--db", "/path/to/rag.lancedb"]
|
|
}
|
|
}
|
|
}
|
|
```
|
|
|
|
Restart `oterm`, then select the `haiku-rag` tools when creating or editing a chat.
|
|
|
|
## Tools
|
|
|
|
| Tool | What it does |
|
|
| ---- | ------------ |
|
|
| `search_documents` | Hybrid search over the database. |
|
|
| `search_documents_by_image` | Search by image. Only offered with a multimodal embedder. |
|
|
| `list_documents` | Titles, URIs and metadata of the stored documents. |
|
|
| `get_document` | A whole document, in reading order. |
|
|
| `get_document_outline` | A document's heading tree, with page numbers. |
|
|
| `get_document_section` | The text of one section of a document. |
|
|
| `execute_code` | Runs a Python program in a sandbox over the selected documents. |
|
|
|
|
The model sees them as `haiku-rag_search_documents` and so on. The haiku.rag [MCP reference](https://ggozad.github.io/haiku.rag/mcp/) documents their arguments.
|
|
|
|
Then ask questions such as:
|
|
|
|
> What does the report say about revenue in the third quarter?
|
|
|
|
> Which of my notes mention the migration, and what did we decide?
|
|
|
|
## Configuration
|
|
|
|
haiku.rag reads `haiku.rag.yaml` from the path in `HAIKU_RAG_CONFIG_PATH`, then the current directory, then its own data directory. The MCP server runs in the directory `oterm` was started from, so either set `cwd` on the server or point it at the file:
|
|
|
|
```json
|
|
{
|
|
"mcpServers": {
|
|
"haiku-rag": {
|
|
"command": "uvx",
|
|
"args": ["haiku-rag", "mcp", "--stdio", "--db", "/path/to/rag.lancedb"],
|
|
"env": {
|
|
"HAIKU_RAG_CONFIG_PATH": "/path/to/haiku.rag.yaml"
|
|
}
|
|
}
|
|
}
|
|
}
|
|
```
|
|
|
|
The embedding model in that configuration must match the one the database was built with. See the haiku.rag [configuration guide](https://ggozad.github.io/haiku.rag/configuration/) for providers, models and reranking.
|