pyproject.toml

[build-system]
requires = ["setuptools>=68", "wheel"]
build-backend = "setuptools.build_meta"

[project]
name = "serviette"
version = "0.1.5"
description = "A universal, no-code RAG server for any vector database, built on the Pathway Live Data Framework."
readme = "README.md"
license = { file = "LICENSE" }
authors = [{ name = "serviette contributors" }]
keywords = ["rag", "pathway", "vector-database", "retrieval", "llm", "embeddings"]
# Pathway requires CPython >= 3.10.
requires-python = ">=3.10"
classifiers = [
    "Programming Language :: Python :: 3",
    "Programming Language :: Python :: 3.10",
    "Programming Language :: Python :: 3.11",
    "Programming Language :: Python :: 3.12",
    "License :: OSI Approved :: MIT License",
    "Operating System :: OS Independent",
]
dependencies = [
    # >=0.32.1: the first release with the vector-store output connectors
    # (duckdb, schema-driven qdrant, pinecone, ...) and lazy xpack imports.
    # Heavy document parsers (docling, unstructured) are opt-in via the
    # serviette extras below.
    "pathway[xpack-llm]>=0.32.1",
    "fastapi>=0.110",
    "uvicorn>=0.27",
    "httpx>=0.27",
    "pydantic>=2.0",
    "pyyaml>=6.0",
    # The default zero-setup vector backend (embedded, pure-wheel).
    "duckdb>=1.0",
    # Explicit, not via pathway[xpack-llm]'s transitive pin: the OpenAI
    # client powers the default /rag LLM and the OpenAI embedder (~6 MB).
    "openai>=1.0",
    # The popular formats must work in a bare `pip install serviette`:
    # pypdf for PDF (tiny, pure-python; the docling extra upgrades it to
    # layout-aware parsing), unstructured for the office family (DOCX, PPTX,
    # XLSX, HTML, EML). Same version pathway's xpack-llm-docs extra pins, so
    # adding serviette[docling] later never conflicts. Heavy stacks stay
    # opt-in: docling (layout/tables), ocr (scans), API keys (audio/video).
    "pypdf>=4.0",
    "unstructured[docx,pptx,xlsx]~=0.18.1",
    # Resolver guard, not a real dependency of ours: unstructured requires
    # an unpinned `numba`, and on py3.12 a clean resolve can backtrack to
    # numba 0.53 whose llvmlite predates py3.10 and fails to build. 0.59 is
    # the first numba with py3.12 wheels.
    "numba>=0.59",
    # Resolver guards for pathway[xpack-llm]: its google-generativeai pin needs
    # protobuf<6, while pathway's unbounded google-cloud-pubsub / google-api-core
    # resolve to releases that need protobuf>=6.33.5. Without these caps pip
    # 25.1-25.3 backtracks through ~30 pubsub releases and dies with
    # resolution-too-deep; older pip wanders for hours.
    "protobuf<6",
    "google-cloud-pubsub<2.40",
    "google-api-core<2.34; python_version < '3.14'",
    # google-api-core>=2.26 on 3.14 wants grpcio-status>=1.75.1, which the
    # xpack-llm extra forbids (<1.72).
    "google-api-core<2.26; python_version >= '3.14'",
]

[project.optional-dependencies]
gdrive = ["google-api-python-client>=2.0", "google-auth>=2.0"]
# The S3 source reads object bytes through a boto3 client. pathway >= 0.33
# no longer depends on boto3 itself, so it is an explicit extra here.
s3 = ["boto3>=1.26"]
sharepoint = ["Office365-REST-Python-Client>=2.5"]
# FTP/SFTP/WebDAV/ZIP and more via the PyFilesystem2 library.
# setuptools<81: PyFilesystem2 imports pkg_resources at runtime; modern venvs
# don't ship it, and setuptools>=81 removed it entirely.
pyfilesystem = ["fs>=2.4", "setuptools<81"]
# Local, free multimodal parsing (the keyless defaults of the parser: section
# pick these up automatically when installed).
docling = ["pathway[xpack-llm-docs]"]  # layout-aware PDF, DOCX via unstructured
# paddlepaddle pinned below 3.3: its PIR+oneDNN executor crashes on some
# CPUs (ConvertPirAttribute2RuntimeAttribute); 3.0.x is solid.
ocr = ["paddleocr>=2.7", "paddlepaddle>=2.6,<3.3", "paddlex[ocr]>=3.0"]
# One extra per vector-DB backend — install only what you use.
pgvector = ["asyncpg>=0.29"]
milvus = ["pymilvus>=2.4"]
qdrant = ["qdrant-client>=1.10"]
chroma = ["chromadb-client>=0.5"]
weaviate = ["weaviate-client>=4.7"]
# <10: pinecone 10 rewrites create_index(dimension=...) into the new
# schema-based request body, which pinecone-local (and our prepare step) do
# not speak yet.
pinecone = ["pinecone>=5.0,<10"]
mongodb = ["pymongo>=4.9"]
# Local, credential-free embeddings (indexer and server side).
local = ["sentence-transformers>=3.0"]
gemini = ["google-generativeai>=0.8"]
# Server-side Bedrock query embedder (the indexer side uses the xpack's
# aioboto3, which pathway[xpack-llm] brings along).
bedrock = ["boto3>=1.26"]
all = [
    "serviette[gdrive,s3,bedrock,pgvector,milvus,qdrant,chroma,weaviate,pinecone,mongodb]",
]
dev = [
    "pytest>=8.0",
    "pytest-asyncio>=0.23",
    "ruff>=0.16,<0.17",
    "mypy>=1.8",
    # Clients for the per-backend integration tests (each test skips when
    # Docker or its client is unavailable).
    "asyncpg>=0.29",
    "pymilvus>=2.4",
    "qdrant-client>=1.10",
    "chromadb-client>=0.5",
    "weaviate-client>=4.7",
    # Fixture builders for the multimodal parser tests (tiny, pure-python).
    "fpdf2>=2.7",
    "python-docx>=1.1",
    # asyncio support is built into pinecone since 9.x (the old [asyncio]
    # extra is gone and only produces resolver warnings).
    "pinecone>=9.0,<10",
    "pymongo>=4.9",
    # Bedrock server-embedder unit tests and the S3 integration test.
    "boto3>=1.26",
    # Embedded Milvus engine for the Milvus integration test (the pgvector
    # integration test instead spins up a Docker container — see docs).
    "milvus-lite>=2.4",
]

[project.scripts]
serviette = "serviette.cli:main"

[project.urls]
Homepage = "https://pathway.com"
Documentation = "https://github.com/pathwaycom/serviette"

[tool.setuptools.packages.find]
include = ["serviette*"]

[tool.setuptools.package-data]
"serviette.frontend" = ["static/*"]
"serviette.demo" = ["corpus/*.md"]

[tool.pytest.ini_options]
asyncio_mode = "auto"
testpaths = ["tests"]
markers = [
    "slow: end-to-end tests that spin up a Pathway runtime per pass",
    "integration: tests that require Docker (real database containers)",
]

[tool.ruff]
line-length = 100
target-version = "py310"

[tool.ruff.lint.per-file-ignores]
# Integration tests import after pytest.importorskip by design.
"tests/test_integration_*.py" = ["E402"]