fleet-memory/hindsight-api/pyproject.toml
Nicolò Boschi 1caf5ec9ee
feat: add jina-mlx reranker provider for Apple Silicon (#542)
* feat: add JinaMLXCrossEncoder for native Apple Silicon reranking

Adds a new `jina-mlx` reranker provider backed by jinaai/jina-reranker-v3-mlx,
a 0.6B multilingual listwise reranker running via the MLX framework on Apple Silicon.
The model is downloaded automatically from HuggingFace Hub on first use.

Benchmarked latencies (Apple Silicon): 1 doc→32ms, 5→45ms, 10→60ms, 20→94ms.
Sub-linear scaling because all docs are ranked in a single forward pass.

- Embeds the MLX reranker implementation (_MLXReranker / _MLPProjector) directly
  in cross_encoder.py with no transformers/PyTorch dependency
- Adds `mlx`, `mlx-lm`, `safetensors` to pyproject.toml optional deps (uv add)
- Updates configuration.md with provider docs and benchmark table

* refactor: import MLXReranker from repo rerank.py instead of duplicating code

Use importlib to load MLXReranker directly from the model repo's own rerank.py
(downloaded via snapshot_download). Also pin exact minimum versions for
mlx>=0.31.0, mlx-lm>=0.31.1, safetensors>=0.6.2 (verified against installed versions).

* refactor: move MLX reranker impl to dedicated jina_mlx_reranker.py

Replaces the importlib hack with a proper module. jina_mlx_reranker.py is
adapted from jinaai/jina-reranker-v3-mlx/rerank.py (CC BY-NC 4.0) with the
source clearly documented at the top of the file.

* docs: simplify jina-mlx reranker docs

* fix: disable GIN fastupdate on source_memory_ids index to prevent deadlocks

GIN fastupdate buffers inserts in a pending list and flushes it with
AccessExclusiveLock when full. Under concurrent test load (8 xdist workers
all running retain_async), two workers can trigger a flush simultaneously
and deadlock. Recreating the index with fastupdate=off eliminates the
flush/lock cycle at the cost of slightly slower individual inserts.

* fix: drop per-bank HNSW indexes after transaction to avoid AccessExclusiveLock deadlock

When deleting a bank, the previous code dropped HNSW indexes inside the
same transaction as the DELETE FROM memory_units. Since DROP INDEX needs
AccessExclusiveLock on the parent table and DELETE holds RowExclusiveLock,
two concurrent bank deletions deadlocked on the same table lock.

Fix: capture internal_id inside the transaction, commit, then drop the
indexes outside the transaction so no row-level locks are held.
2026-03-11 15:15:58 +01:00

190 lines
5.8 KiB
TOML

[build-system]
requires = ["hatchling"]
build-backend = "hatchling.build"
[project]
name = "hindsight-api"
version = "0.4.17"
description = "Hindsight: Agent Memory That Works Like Human Memory"
readme = "README.md"
requires-python = ">=3.11"
dependencies = [
"asyncpg>=0.29.0",
"python-dotenv>=1.0.0",
"openai>=1.0.0",
"pydantic>=2.0.0",
"rich>=13.0.0",
"langchain-text-splitters>=0.3.0",
"fastapi[standard]>=0.120.3",
"uvicorn>=0.38.0",
"wsproto>=1.0.0",
"sqlalchemy>=2.0.44",
"alembic>=1.17.1",
"pgvector>=0.4.1",
"greenlet>=3.2.4",
"psycopg2-binary>=2.9.11",
"tiktoken>=0.12.0",
"httpx>=0.27.0",
"PyJWT[crypto]>=2.8.0",
"fastmcp>=2.14.0", # CVE-2025-66416
"pg0-embedded>=0.11.0",
"python-dateutil>=2.8.0",
"opentelemetry-api>=1.20.0",
"opentelemetry-sdk>=1.20.0",
"opentelemetry-instrumentation-fastapi>=0.41b0",
"opentelemetry-exporter-prometheus>=0.41b0",
"opentelemetry-exporter-otlp-proto-http>=1.20.0",
"opentelemetry-semantic-conventions>=0.41b0",
"dateparser>=1.2.2",
"google-genai>=1.0.0",
"google-auth>=2.0.0",
"anthropic>=0.40.0",
"typer>=0.9.0",
"cohere>=5.0.0",
"flashrank>=0.2.0",
"litellm>=1.0.0",
"markitdown[pdf,docx,pptx,xlsx,xls]>=0.1.4", # File to markdown conversion
"obstore>=0.4.0", # S3/GCS/Azure object storage client (Rust-backed)
# Local ML models for embeddings/reranking - can be excluded in Docker with INCLUDE_LOCAL_MODELS=false
"sentence-transformers>=3.3.0",
"transformers>=4.53.0", # Security fixes for ReDoS vulnerabilities
"torch>=2.6.0", # CVE fix for remote code execution
"uvloop>=0.22.1",
# Transitive dependency security fixes
"pyasn1>=0.6.2", # DoS vulnerability fix
"urllib3>=2.6.3", # Decompression-bomb safeguards bypass fix
"langchain-core>=1.2.11", # Serialization injection + SSRF vulnerability fix
"langsmith>=0.6.3", # SSRF via tracing header injection fix
"protobuf>=6.33.5", # JSON recursion depth bypass fix
"pillow>=12.1.1", # Out-of-bounds write in PSD image loading fix
"cryptography>=46.0.5", # Subgroup attack vulnerability fix
"filelock>=3.20.1", # TOCTOU race condition fix
"authlib>=1.6.6", # Account takeover vulnerability fix
"aiohttp>=3.13.3", # Multiple DoS vulnerabilities
"claude-agent-sdk>=0.1.27",
"einops>=0.8.2",
"mlx>=0.31.0",
"mlx-lm>=0.31.1",
"safetensors>=0.6.2",
]
[project.optional-dependencies]
test = [
"pytest>=7.0.0",
"pytest-asyncio>=0.21.0",
"pytest-timeout>=2.4.0",
"pytest-xdist>=3.0.0",
"filelock>=3.20.1", # TOCTOU race condition fix
"testcontainers>=4.0.0",
]
[project.scripts]
hindsight-api = "hindsight_api.main:main"
hindsight-worker = "hindsight_api.worker.main:main"
hindsight-local-mcp = "hindsight_api.mcp_local:main"
hindsight-admin = "hindsight_api.admin.cli:main"
[tool.hatch.build.targets.wheel]
packages = ["hindsight_api"]
[tool.hatch.build.targets.wheel.sources]
"hindsight_api" = "hindsight_api"
[tool.hatch.build.targets.sdist]
include = [
"hindsight_api/**/*",
]
[tool.hatch.build]
include = [
"hindsight_api/**/*.py",
"hindsight_api/alembic/**/*",
]
[tool.pytest.ini_options]
log_cli = true
log_cli_level = "INFO"
log_cli_format = "%(asctime)s - %(levelname)s - %(name)s - %(message)s"
log_cli_date_format = "%Y-%m-%d %H:%M:%S"
addopts = "--timeout 300 -n 8 --dist loadgroup --durations=10 -v"
asyncio_mode = "auto"
asyncio_default_fixture_loop_scope = "function"
log_auto_indent = true
filterwarnings = [
"ignore:The @wait_container_is_ready decorator is deprecated:DeprecationWarning",
"ignore::RuntimeWarning:asyncio",
]
[dependency-groups]
dev = [
"pytest>=9.0.0",
"pytest-asyncio>=1.3.0",
"pytest-timeout>=2.4.0",
"pytest-xdist>=3.8.0",
"python-dotenv>=1.2.1",
"filelock>=3.20.1", # TOCTOU race condition fix
"ruff>=0.8.0",
"ty>=0.0.1",
"testcontainers>=4.0.0",
]
[tool.ruff]
line-length = 120
target-version = "py311"
exclude = [
"tests/",
"**/tests/",
]
[tool.ruff.lint]
select = [
"E", # pycodestyle errors
"W", # pycodestyle warnings
"F", # Pyflakes
"I", # isort
]
ignore = [
"E501", # line too long (handled by formatter)
"E402", # module import not at top of file
"F401", # unused import (too noisy during development)
"F841", # unused variable (too noisy during development)
"F811", # redefined while unused
"F821", # undefined name (forward references in type hints)
]
[tool.ruff.lint.isort]
known-third-party = ["alembic"]
[tool.ruff.format]
quote-style = "double"
indent-style = "space"
[tool.uv]
# Allow uv to search all configured indexes for packages, not just the first one
# This prevents dependency resolution failures when using pytorch index + PyPI
index-strategy = "unsafe-best-match"
[tool.ty]
# Type checking configuration
# ty is an extremely fast Python type checker from Astral (same team as ruff/uv)
[tool.ty.environment]
python-version = "3.11"
[tool.ty.src]
exclude = [
"tests/",
"hindsight_api/alembic/",
]
[tool.ty.rules]
# Disable noisy rules while keeping important ones
invalid-argument-type = "ignore" # False positives with **kwargs patterns
invalid-return-type = "ignore" # Often intentional in async code
invalid-parameter-default = "ignore" # Optional params with None default
possibly-missing-attribute = "ignore" # Common with Optional types
invalid-raise = "ignore" # False positives with exception tracking
call-non-callable = "ignore" # False positives with Optional types
invalid-key = "ignore" # Pydantic ConfigDict not understood
invalid-method-override = "ignore" # Intentional signature differences
unresolved-reference = "ignore" # Forward references not always resolved