jambashrimp 0.1.0

JambaShrimp (jash): playful CLI to chat with PDFs or crawled docs using Jamba Reasoning 3B

JambaShrimp (jash)

Playful command-line interface for chatting with any PDF or crawled documentation site using AI21's Jamba Reasoning 3B (GGUF build). The CLI can:

• load entire PDFs directly into the prompt for ad-hoc question answering
• crawl documentation sites (recursively) and ground answers with a local FAISS index (RAG)
• cache indexes per URL so you can come back later—even offline—and keep chatting

The product name is JambaShrimp, and its CLI entry point is the shortened jash.

Requirements

• uv for environment management
• macOS with Metal (recommended) or CPU fallback
• Optional HF_TOKEN environment variable if your Hugging Face account is required for downloads

Install with pipx

Publishing the package (named jambashrimp) to PyPI lets anyone install the CLI with an isolated virtual environment:

pipx install jambashrimp

# run it
jash --help

# download the sample PDF + model
jash-setup

While developing locally you can point pipx directly at the repo:

pipx install --spec . jambashrimp

Assets are stored under ~/.jambashrimp/assets when the package is installed globally. You can override --model/--pdf if you keep files elsewhere.

Dev Setup

uv sync
uv run scripts/setup_assets.py     # downloads the default OECD PDF + GGUF model
uv run python -m jamba_cli.cli

Pass --pdf or --model to point at different files. Run uv run jamba-chat --help to see all knobs (context window, temperature, GPU layers, etc.).

CLI Shortcuts

• /quit or /exit – leave the chat
• /history – print the running conversation
• /reload – reload the PDF from disk (useful when editing)
• /help – show the command list

By default the CLI hides <think> traces but still streams answers token-by-token. Useful switches:

• --history-turns N – keep N previous Q/A pairs inside the prompt (defaults to 0 to maximize context for the document itself).
• --cache-prompt – reuse the KV cache between turns if you really need faster follow ups. Leave it off (default) if you encounter llama_decode returned -1; clearing the cache between turns prevents those context overflows.
• --show-thinking / --no-stream – reveal hidden reasoning traces or print whole answers at once.