# Copyright 1999-2026 Gentoo Authors # Distributed under the terms of the GNU General Public License v2 EAPI=8 DISTUTILS_USE_PEP517=setuptools PYTHON_COMPAT=( python3_{12..14} ) DISTUTILS_SINGLE_IMPL=1 inherit distutils-r1 DESCRIPTION="Lightweight agent framework for the edge and AMD Ryzen AI PCs" HOMEPAGE="https://github.com/amd/gaia" SRC_URI="https://github.com/amd/gaia/archive/refs/tags/v${PV}.tar.gz -> ${P}.gh.tar.gz" S="${WORKDIR}/gaia-${PV}" LICENSE="MIT" SLOT="0" KEYWORDS="~amd64" IUSE="+api audio +mcp eval image talk ui" # ui implies api: upstream's ui extra restates fastapi/uvicorn/python- # multipart on top of its own RAG deps. REQUIRED_USE="ui? ( api )" # Upstream pytest config marks tests as needing a live Lemonade server, # Docker, the Gmail API, and other integration targets that aren't # reachable from a sandboxed build. RESTRICT until someone wants to # split out a unit-only subset. RESTRICT="test" # Core install_requires per setup.py (verified against v0.21.2 sdist # 2026-06-19). # Lemonade Server is the recommended OpenAI-compatible LLM backend but # isn't a hard dep — gaia speaks any OpenAI-compatible endpoint, so the # RDEPEND only mentions the bundled Python deps. See pkg_postinst. # # python-multipart — base install_requires since v0.20.1: the base # `gaia-mcp` console_script parses multipart uploads via python_multipart # at import time, so a plain install needs it even without USE=api. # Carried in base and dropped from api? as redundant. still base in # v0.21.2, verified 2026-06-19. # # keyring — v0.21.0 promotes it from the ui/api extras into the base # install_requires (upstream #1621): gaia.connectors.{store,mcp_server} # both `import keyring` at module load and `gaia connectors` is a base CLI # command, so a plain install needs it. Carried in base (>=24,<26 per # upstream's supply-chain pin) and dropped from ui? as redundant. verified # 2026-06-19 against v0.21.2. # # tavily-python — v0.21.2 adds it to base install_requires (>=0.5.0), but # it's a SOFT dep: gaia/web/tavily.py guards `from tavily import ...` in a # try/except (TAVILY_SDK_AVAILABLE) and falls back to DuckDuckGo when the # SDK is absent. Not packaged in ::gentoo or this overlay, so omitted from # RDEPEND — web search degrades gracefully without it. verified 2026-06-19. # # audio? — gaia code only `import torch`s (gaia/audio/whisper_asr.py); # never imports torchvision OR torchaudio (re-grepped src/ for v0.21.2, # 2026-06-19). Upstream's `audio` extra caps torch<2.13 because of # old-era openai-whisper transitive deps; verified 2026-06-04 the cap is # stale (current openai-whisper has unbounded torch and no torchvision # dep), so we ship without the cap and without the unused torchvision/ # torchaudio. RDEPEND on sci-ml/pytorch alone. # # ui? — upstream 0.20.0 added python-pptx>=0.6.21 (PPTX ingestion in # RAG). gaia.rag.{sdk,pptx_utils} both `from pptx import ...` lazily, so # the missing dep only surfaces if a user actually ingests a .pptx # file. dev-python/python-pptx isn't packaged in ::gentoo or this # overlay yet; revisit once it lands. still present in v0.21.2, verified # 2026-06-19. # # httpx (ui?) — a hard upstream requirement: the `ui` extra declares # httpx>=0.27.0 (setup.py) and 9 src/gaia modules import it (ui/server.py, # ui/tunnel.py, ui/routers/*, agents/base/agent.py, ...). httpx is # ::gentoo-deprecated (2026-04-01: upstream stopped accepting bug reports) # but is still in-tree with no drop-in replacement, so the DeprecatedDep # warning is knowingly accepted, not fixable by removal. Revisit if/when # ::gentoo last-rites httpx. verified 2026-06-20. # # eval? — v0.21.2 adds tiktoken>=0.7.0,<1 to the eval extra (token-cost # accounting in gaia/eval/tool_cost.py; imported lazily with a char-count # fallback when absent). dev-python/tiktoken is packaged, carried uncapped. # Upstream also caps numpy>=2.0,<2.3.0 here, but ::gentoo ships numpy 2.4+ # by default so that cap is left off (same posture as the stale torch cap # above) — would otherwise make USE=eval unsolvable. verified 2026-06-19. # # talk? — v0.21.0 adds `pip` to the talk extra: the Kokoro/misaki TTS # path downloads its spaCy model at runtime via pip. still in v0.21.2, # verified 2026-06-19. RDEPEND=" ${PYTHON_DEPS} sci-ml/accelerate[${PYTHON_SINGLE_USEDEP}] sci-ml/transformers[${PYTHON_SINGLE_USEDEP}] $(python_gen_cond_dep ' dev-python/aiohttp[${PYTHON_USEDEP}] dev-python/beautifulsoup4[${PYTHON_USEDEP}] >=dev-python/keyring-24.0.0[${PYTHON_USEDEP}] =dev-python/pillow-9.0.0[${PYTHON_USEDEP}] >=dev-python/pydantic-2.9.2[${PYTHON_USEDEP}] dev-python/python-dotenv[${PYTHON_USEDEP}] >=dev-python/python-multipart-0.0.9[${PYTHON_USEDEP}] dev-python/requests[${PYTHON_USEDEP}] dev-python/rich[${PYTHON_USEDEP}] >=dev-python/watchdog-2.1.0[${PYTHON_USEDEP}] api? ( >=dev-python/fastapi-0.115.0[${PYTHON_USEDEP}] >=dev-python/uvicorn-0.32.0[${PYTHON_USEDEP}] ) image? ( dev-python/term-image[${PYTHON_USEDEP}] ) talk? ( dev-python/pip[${PYTHON_USEDEP}] dev-python/sounddevice[${PYTHON_USEDEP}] dev-python/soundfile[${PYTHON_USEDEP}] dev-python/psutil[${PYTHON_USEDEP}] ) ui? ( >=dev-python/httpx-0.27.0[${PYTHON_USEDEP}] >=dev-python/psutil-5.9.0[${PYTHON_USEDEP}] dev-python/PyMuPDF[${PYTHON_USEDEP}] dev-python/pypdf[${PYTHON_USEDEP}] sci-ml/safetensors[${PYTHON_USEDEP}] ) mcp? ( >=dev-python/mcp-1.1.0[${PYTHON_USEDEP}] dev-python/starlette[${PYTHON_USEDEP}] dev-python/uvicorn[${PYTHON_USEDEP}] ) eval? ( dev-python/anthropic[${PYTHON_USEDEP}] dev-python/numpy[${PYTHON_USEDEP}] dev-python/pypdf[${PYTHON_USEDEP}] dev-python/reportlab[${PYTHON_USEDEP}] >=dev-python/scikit-learn-1.5.0[${PYTHON_USEDEP}] >=dev-python/tiktoken-0.7.0[${PYTHON_USEDEP}] ) ') audio? ( sci-ml/pytorch[${PYTHON_SINGLE_USEDEP}] ) talk? ( dev-python/kokoro[${PYTHON_SINGLE_USEDEP}] dev-python/openai-whisper[${PYTHON_SINGLE_USEDEP}] ) ui? ( sci-libs/faiss[python] sci-ml/sentence-transformers[${PYTHON_SINGLE_USEDEP}] ) " DEPEND="${RDEPEND}" BDEPEND="${PYTHON_DEPS}" pkg_postinst() { elog "GAIA is an LLM-agent framework. It speaks any OpenAI-compatible" elog "endpoint; the AMD-recommended local backend is Lemonade Server" elog "(sci-ml/lemonade in this overlay), which runs models on Ryzen AI" elog "hardware (NPU + iGPU). Point gaia at a server with:" elog "" elog " export OPENAI_BASE_URL=http://localhost:8000/api/v1" elog "" elog "Extras supported via USE flags:" elog " audio — sci-ml/pytorch (gaia code doesn't touch torchvision/" elog " torchaudio despite upstream's audio extra listing them)" elog " image — dev-python/term-image" elog " talk — dev-python/openai-whisper + dev-python/sounddevice +" elog " dev-python/kokoro (full upstream parity)." elog " ui — full RAG-over-PDFs web frontend (faiss + sentence-" elog " transformers + PyMuPDF + pypdf + safetensors);" elog " implies +api" elog "" elog "Extras still not built (deps not all in tree):" elog " blender — bpy (Blender Python module — heavy)" elog "" elog "Web search uses the Tavily SDK (dev-python/tavily-python, not" elog "packaged); without it gaia falls back to DuckDuckGo automatically." elog "" elog "Use the upstream pip install if you need an extra flavour we" elog "haven't packaged yet." }