remex-cli 2.0.0

Local-first RAG library — ingest files and SQLite, query semantically, pipe results into any AI agent

Remex

Stop searching. Start asking.

Local-first AI for the documents you already have.

Remex ingests PDFs, Word docs, Markdown, spreadsheets, code, web pages, e-books, and SQLite databases into a local vector index, then lets you ask questions in plain language and get answers cited from your own files. Everything — embedding, storage, retrieval — runs on your machine. The AI provider is optional and you bring your own (Anthropic, OpenAI, Ollama, or any OpenAI-compatible endpoint).

---

Install

🖥️ Remex Studio Native desktop app for Windows. Zero terminal. ⬇ Download the latest release [!WARNING] Windows SmartScreen may flag the installer because Remex isn't code-signed with a paid certificate. The source is in this repo — audit freely. Click More info → Run anyway to proceed.

⌨️ Python CLI pip install remex-cli # core (7 formats) pip install "remex-cli[formats]" # +.pptx .xlsx .epub .html .odt pip install "remex-cli[ai]" # +Anthropic & OpenAI SDKs pip install "remex-cli[sentence]" # +NLTK sentence chunking pip install "remex-cli[all]" # everything

---

_{Ask a question, get a synthesised answer grounded in your own files — with the source chunks and similarity scores one click away.}

---

Features

🔍 Search & answer Semantic search across one or many collections AI Answer — synthesised, grounded, sources cited AI personas — save voices (Researcher, Code Q&A, …) and switch from the Query pane Low-confidence banner when no chunk scores above 0.70 Source filter to narrow before searching Searchable query history Chunk viewer with keyboard navigation Export to JSON · CSV · Markdown · BibTeX · RIS · CSL-JSON · Obsidian vault

📥 Ingest 12 file formats: .pdf .docx .md .txt .csv .json .jsonl plus .html .pptx .xlsx .epub .odt SQLite ingest — single table or a custom SQL query with JOINs Recursive / sentence / word chunking Incremental — SHA-256 hash check, only changed rows / files re-processed Auto-ingest folders and SQLite files on change (debounced)

⚙️ Manage & customise All embedding models — MiniLM, BGE, multilingual, Nomic long-context, plus any HuggingFace / FastEmbed / Ollama model by name Custom OpenAI-compatible providers for corporate gateways (Azure OpenAI, LiteLLM, vLLM) Custom embedding presets Per-project defaults for chunk size, overlap, embedding model, max results, min score Collections manager — rename, describe, purge, bulk-delete, re-ingest Light · dark · auto themes, 16 accent colours, 3 home backgrounds Keyboard-first — press ? for shortcuts

Remex is free and open-source. Every feature ships in the box — no tiers, no license keys, no payment required.

---

Privacy & data flow

Remex is local-first. Here's exactly what touches the network:

Operation	Where it runs	Network?
File / SQLite extraction & chunking	🖥 Local	✗ None
Default embedding (all-MiniLM-L6-v2 or any FastEmbed model)	🖥 Local (ONNX on CPU)	✗ None
Vector storage (ChromaDB)	🖥 Local (./remex_db/)	✗ None
Vector search (AI Answer off)	🖥 Local	✗ None
AI Answer (on)	☁ Your AI provider	✓ Question + top-N retrieved chunks sent to the endpoint you configured
OpenAI embedding model (opt-in)	☁ OpenAI	✓ Document text sent to OpenAI during ingest
API key storage	🔐 OS keyring	✗ Never on disk in plaintext, never transmitted except as the auth header on requests you initiate
Sidecar HTTP listener	127.0.0.1 loopback only	✗ Not reachable from the LAN

[!IMPORTANT] Remex itself has no analytics, no telemetry, and no controlled servers in the request path. The only outbound traffic is to the AI provider you yourself configured — and only when AI Answer is enabled or you submit a query in AI mode.

For zero data egress, point AI Answer at Ollama (local) or a self-hosted vLLM/TGI via a Custom Provider.

---

CLI quick start

# Scaffold a project (creates ./docs, remex.toml, .gitignore entries)
remex init

# Ingest a folder of documents
remex ingest ./docs

# Semantic search
remex query "how does authentication work?"

# AI-synthesised answer (requires ANTHROPIC_API_KEY, OPENAI_API_KEY, or a running Ollama)
remex query "how does authentication work?" --ai

Command reference

Command	Description
remex init [path]	Scaffold docs/, remex.toml, .gitignore entries
remex ingest <dir>	Ingest files from a directory
remex ingest-sqlite <db>	Ingest rows from a SQLite table or SQL query
remex query <text>	Semantic search; add --ai for an AI answer
remex sources	List ingested source paths
remex stats	Show chunk and source counts
remex delete-source <path>	Remove all chunks for a source
remex purge	Remove chunks whose source file no longer exists
remex reset	Wipe a collection
remex list-collections	List collections in a database
remex serve	Start the FastAPI sidecar on localhost:8000

remex <command> --help    # full option reference for any command

Use as a library

from remex import ingest, query

# Ingest a folder
result = ingest("./docs", collection_name="my-kb")
print(f"{result.chunks_stored} chunks stored")

# Search
results = query("how does auth work?", collection_name="my-kb")
for r in results:
    print(f"[{r.score:.3f}] {r.source}  →  {r.text[:120]}")

---

Configuration

Drop a remex.toml in your project root (or run remex init):

[remex]
db              = "./remex_db"
collection      = "my-kb"
embedding_model = "all-MiniLM-L6-v2"

# chunk_size     = 768          # characters per chunk (512–1024 works well)
# overlap        = 150          # ~20% overlap preserves context at boundaries
# min_chunk_size = 50           # discard chunks shorter than this
# chunking       = "recursive"  # "recursive" (default) | "sentence" | "word"

CLI flags always override remex.toml. Studio writes per-project defaults via Settings → General → Project defaults when a project is open.

---

Embedding models

Preset	Model	Size	Notes
Light (default)	all-MiniLM-L6-v2	22 MB	Fast, good accuracy
Balanced	intfloat/e5-base-v2	438 MB	Better retrieval quality
Multilingual	paraphrase-multilingual-MiniLM-L12-v2	470 MB	50+ languages
Large	BAAI/bge-large-en-v1.5	1.3 GB	Best English accuracy
E5 Large	intfloat/e5-large-v2	1.3 GB	Strong retrieval benchmark
Long ctx	nomic-ai/nomic-embed-text-v1.5	547 MB	8 192-token context

Any SBERT, HuggingFace, or Ollama embedding model works — type the model name into the picker. Add your own to Settings → AI & Server → Embedding presets.

[!NOTE] Score interpretation. ChromaDB returns L2² distance; Remex converts it to score = 1 / (1 + distance) so values lie in (0, 1]. For normalised embeddings (BGE, MiniLM, Nomic, OpenAI): 0.75+ is near-paraphrase, 0.65–0.75 is strongly relevant, 0.50–0.55 is borderline noise.

---

Under the hood

┌────────────────┐   Tauri IPC    ┌────────────────────┐   HTTP/127.0.0.1   ┌──────────────────┐
│  Studio (TS)   │ ─────────────► │  Rust shell        │ ────────────────► │  Python sidecar  │
│  React + Vite  │ ◄───────────── │  (lib.rs)          │ ◄──────────────── │  FastAPI         │
└────────────────┘                └────────────────────┘                    └────────┬─────────┘
                                                                                    │
                                                                                    ▼
                                                                          ┌──────────────────┐
                                                                          │  ChromaDB        │
                                                                          │  FastEmbed/ONNX  │
                                                                          │  ./remex_db/     │
                                                                          └──────────────────┘

• Studio — Tauri 2, React 18, Zustand, react-query, shadcn/Radix UI
• Sidecar — FastAPI on 127.0.0.1, uv-managed venv in %APPDATA%\com.remex.studio\
• Indexing — ChromaDB persistent store, FastEmbed ONNX runtime, recursive / sentence / word chunker
• Secrets — OS keyring via the keyring crate (Windows Credential Manager · macOS Keychain · libsecret)

---

Build from source

[!TIP] Studio requires Rust, Node.js 20+, and the Tauri prerequisites for your OS.

# Python CLI
pip install -e ".[dev]"
pytest

# Studio (dev server with hot-reload)
cd studio
npm install
npm run tauri dev

# Studio (production build)
npm run tauri build

See studio/README.md for the full Studio build guide.

---

Changelog · Contributing · Licensing · Issues · GitHub

_{Python CLI: Apache-2.0 · Studio (v1.3.0+): FSL-1.1-MIT — see LICENSES.md}