forktex-core 2.1.0

Shared Python substrate for ForkTex services — async Postgres, Redis, durable execution, encryption, S3/MinIO, background jobs, vector search, tenant-defined Grid schemas, and structured logging

Forktex Core

The shared Python substrate for Forktex services. Pick-and-choose extras across five levels. One install. Zero framework lock-in.

pip install forktex-core                 # log · database · cache · queue · grid · flow
pip install forktex-core[vault]          # + Fernet encryption
pip install forktex-core[storage]        # + S3/MinIO blob storage
pip install forktex-core[vector]         # + Qdrant vector search
pip install forktex-core[all]            # everything

---

Architecture

The five-level architecture below is the single source of truth for what extras exist, how they relate, and which infra each requires. Tables are generated from forktex_core/catalog/catalog.json — refresh with make catalog-update; make catalog-check runs in CI.

Levels overview

Level	Name	Description	Extras
0	primitives	Zero-dep cross-cutting utilities. Always pulled.	log · error · types
1	tech_adapters	Raw client per infrastructure service. Exposes connection pool + low-level ops.	postgres · redis · qdrant · minio · mongo
2	role_facades	Role-named abstractions over tech adapters. Expose Database() / Cache() / Queue() / Vector() / Storage() / Store() / Vault() / Graph() user-facing classes.	database · cache · queue · vector · storage · store · vault · graph
3	substrate_facades	Substrate user-facing pillars composing level 0–2.	grid · space · flow
4	bootstraps	Process-level runtime wiring (FastAPI factory, arq worker bootstrap).	api · worker

Level 0 — Primitives

Zero-dep cross-cutting utilities. Always pulled.

Extra	Role	Status
	Structured logging protocol + request-id context propagation.	✅ shipped
	AppError hierarchy + envelope shape + http exception mapper.	✅ shipped
	Base Pydantic models, frozen value objects, JSON-Schema conventions.	✅ shipped

Level 1 — Tech adapters

Raw client per infrastructure service. Exposes connection pool + low-level ops.

Planned for the 1.x line — see GitHub issues for status.

Level 2 — Role facades

Role-named abstractions over tech adapters. Expose Database() / Cache() / Queue() / Vector() / Storage() / Store() / Vault() / Graph() user-facing classes.

Extra	Role	Depends on	Status
	Relational/ACID structured-state. Database() over [postgres] adapter: BaseDBModel, AuditMixin, get_session, advisory locks, migrations.	log, error, types, postgres	✅ shipped
	Key-value cache. Cache() over [redis] adapter: get/set, decorators, prefix invalidation.	log, error, types, redis	✅ shipped
	Async job queue. Queue() over [redis] adapter: arq tasks, enqueue, JobCtx.	log, error, types, redis	✅ shipped
	Vector store + similarity search. Vector() over [qdrant] adapter: CollectionHandle, VectorPoint, search.	log, error, types, qdrant	✅ shipped
	Object/blob storage. Storage() over [minio] adapter: presigned URLs, multipart upload.	log, error, types, minio	✅ shipped
	Crypto-at-rest helpers (Fernet over [database] state).	log, error, types, database	✅ shipped
	In-memory multi-edge typed-graph algebra. Graph() with BFS/DFS/closure/cycle, deterministic edge IDs, JSON export. No infra dep.	log, error, types	✅ shipped

Level 3 — Substrate facades

Substrate user-facing pillars composing level 0–2.

Extra	Role	Depends on	Lazy imports	Status
	Tabular state pillar — typed cells, rows, cell-pinned edges, recursive-CTE traversal, JSON-Schema import/export. Pure-tabular: no VECTOR/FILE handlers.	log, error, types, database	—	✅ shipped
	Multi-grid bundle + rich-content wrapper (VECTOR/FILE field types) + cross-grid traversal + SyncSourceConfig contract.	log, error, types, grid, graph	vector, storage	✅ shipped
	Durable workflow execution. Tasks, runs, step events, scheduled runs.	log, error, types, database, queue, cache	—	✅ shipped

Level 4 — Bootstraps

Process-level runtime wiring (FastAPI factory, arq worker bootstrap).

Extra	Role	Required	Optional for consumer	Status
	FastAPI factory: standard middleware stack, lifespan, health probes, error envelope.	log, error, types	grid, space, database, cache, vault	✅ shipped
	arq worker bootstrap: lifecycle, signal handling, optional flow-driver wiring.	log, error, types, queue	database, grid, space, flow	✅ shipped

Pick & choose

Consumer wants…	forktex_core extras	Infra services
Pure tabular registers (basic field types only)	grid	postgres
Tabular registers + vectors (VECTOR field added)	grid, space, vector	postgres, qdrant
Tabular registers + files (FILE field added)	grid, space, storage	minio, postgres
Multi-grid bundle with VECTOR + FILE	grid, space, vector, storage	minio, postgres, qdrant
In-memory graph analysis only	graph	(none)
Just durable workflows	flow	postgres, redis
API server, no DB	api	(none)
API server with grid CRUD	api, grid	postgres
API server with rich content + middleware	api, grid, space, vector, storage, cache	minio, postgres, qdrant, redis
Background worker, pure compute	worker	redis
Background worker with flow runs	worker, flow	postgres, redis
Background worker with grid + flow + vector embedding	worker, flow, grid, space, vector	postgres, qdrant, redis

Filesystem

forktex_core/
│  ── Level 0: primitives ──
│   ├── log/  [log] → —  ✅ shipped
│   │       Structured logging protocol + request-id context propagation.
│   ├── error/  [error] → —  ✅ shipped
│   │       AppError hierarchy + envelope shape + http exception mapper.
│   ├── types/  [types] → —  ✅ shipped
│   │       Base Pydantic models, frozen value objects, JSON-Schema conventions.
│
│  ── Level 2: role_facades ──
│   ├── database/  [database] → —  ✅ shipped
│   │       Relational/ACID structured-state. Database() over [postgres] adapter: BaseDBModel, AuditMixin, get_session, advisory locks, migrations.
│   ├── cache/  [cache] → —  ✅ shipped
│   │       Key-value cache. Cache() over [redis] adapter: get/set, decorators, prefix invalidation.
│   ├── queue/  [queue] → —  ✅ shipped
│   │       Async job queue. Queue() over [redis] adapter: arq tasks, enqueue, JobCtx.
│   ├── vector/  [vector] → —  ✅ shipped
│   │       Vector store + similarity search. Vector() over [qdrant] adapter: CollectionHandle, VectorPoint, search.
│   ├── storage/  [storage] → —  ✅ shipped
│   │       Object/blob storage. Storage() over [minio] adapter: presigned URLs, multipart upload.
│   ├── vault/  [vault] → —  ✅ shipped
│   │       Crypto-at-rest helpers (Fernet over [database] state).
│   ├── graph/  [graph] → —  ✅ shipped
│   │       In-memory multi-edge typed-graph algebra. Graph() with BFS/DFS/closure/cycle, deterministic edge IDs, JSON export. No infra dep.
│
│  ── Level 3: substrate_facades ──
│   ├── grid/  [grid] → —  ✅ shipped
│   │       Tabular state pillar — typed cells, rows, cell-pinned edges, recursive-CTE traversal, JSON-Schema import/export. Pure-tabular: no VECTOR/FILE handlers.
│   ├── space/  [space] → —  ✅ shipped
│   │       Multi-grid bundle + rich-content wrapper (VECTOR/FILE field types) + cross-grid traversal + SyncSourceConfig contract.
│   ├── flow/  [flow] → —  ✅ shipped
│   │       Durable workflow execution. Tasks, runs, step events, scheduled runs.
│
│  ── Level 4: bootstraps ──
│   ├── api/  [api] → —  ✅ shipped
│   │       FastAPI factory: standard middleware stack, lifespan, health probes, error envelope.
│   ├── worker/  [worker] → —  ✅ shipped
│   │       arq worker bootstrap: lifecycle, signal handling, optional flow-driver wiring.
│

---

Usage

log — set up first, before anything else

from forktex_core.log import setup_logging, get_logger, TraceIDMiddleware

setup_logging(service="my-service")       # JSON to stdout, INFO
# setup_logging(service="my-service", debug=True)  # human-readable, DEBUG

log = get_logger(__name__)
log.info("starting up")

# FastAPI: add middleware so every request gets a trace_id automatically
app.add_middleware(TraceIDMiddleware)

database — Postgres connection + ORM

from forktex_core.database import init_engine, get_session, BaseDBModel, NamespacedMixin, AuditMixin
import sqlalchemy as sa
from sqlalchemy.orm import Mapped, mapped_column
import uuid

init_engine("postgresql+asyncpg://user:pass@host/db", pool_size=10)

class Invoice(BaseDBModel, NamespacedMixin, AuditMixin):
    __tablename__ = "invoice"
    id: Mapped[uuid.UUID] = mapped_column(primary_key=True, default=uuid.uuid4)
    amount: Mapped[int] = mapped_column(sa.Integer)

async with get_session() as session:          # auto-commit / rollback
    session.add(Invoice(namespace=str(org_id), amount=100))

cache — Redis

from forktex_core.cache import init, cached

await init("redis://localhost:6379/0")

@cached(ttl=300)
async def get_org(org_id: str) -> dict: ...

# Pair with structured log context (lives in forktex_core.log)
from forktex_core.log import async_log_context
async with async_log_context(org_id=str(org_id)):
    log.info("processing")    # → {..."org_id": "org-xyz"}

flow — durable workflows

from forktex_core.flow import Flow, step

flow = Flow(database_url="postgresql+asyncpg://...")
await flow.init()

@step
async def send_welcome(ctx, state): ...

@flow.pipeline("onboarding.user", version=1)
class UserOnboarding:
    steps = [send_welcome, create_workspace]

instance = await flow.run("onboarding.user", state={"email": "[e-mail]"})
await instance.wait(timeout=60)

vault — encryption at rest

from forktex_core.vault import Vault, EncryptedJSON
import os

vault = Vault(kek=os.environ["FTX_KEK"])

class Provider(BaseDBModel):
    __tablename__ = "provider"
    credentials: Mapped[bytes] = mapped_column(EncryptedJSON(vault))

provider.credentials = {"api_key": "sk-..."}   # transparent encrypt/decrypt

storage — S3/MinIO

from forktex_core.storage import register, get_client

# Register once at startup (supports multiple buckets)
register("docs", url="http://minio:9000", bucket="documents",
         access_key=KEY, secret_key=SECRET)

client = get_client("docs")
await client.upload("invoices/abc.pdf", pdf_bytes, content_type="application/pdf")
url = await client.presign("invoices/abc.pdf", expires_in=3600)

# Actor uploads directly to MinIO — no auth header needed, signature is in the URL
put_url = await client.presign("uploads/photo.jpg", method="put_object",
                               content_type="image/jpeg", expires_in=900)

queue — background jobs

from forktex_core.queue import task, init, enqueue, make_worker, JobCtx

await init("redis://localhost:6379/1")

@task(retries=2, timeout=120)
async def process_document(ctx: JobCtx, doc_id: str) -> None:
    ...

job_id = await enqueue(process_document, str(doc.id))

# Worker entrypoint (run separately)
WorkerSettings = make_worker("redis://localhost:6379/1")

vector — semantic search

from forktex_core.vector import Vector, VectorPoint, SearchQuery

vector = Vector(qdrant_url="http://qdrant:6333")
coll = vector.collection(f"org-{org_id}--knowledge")   # use -- not : as separator
await coll.create(dim=1536)

await coll.upsert([VectorPoint(id=1, vector=embed(text), payload={"text": text})])

hits = await coll.search(SearchQuery(vector=embed(query)).limit(10).using("hybrid"))
for h in hits:
    print(h.score, h.payload["text"])

grid — runtime tabular schemas

from forktex_core.grid import EntityCreate, EntityMode, FieldCreate, FieldDataType, Grid
from forktex_core.database import get_session

async with get_session() as session:
    contacts = await Grid.declare(
        session,
        namespace=str(org_id),
        entity=EntityCreate(slug="contacts", label="Contacts", mode=EntityMode.VIRTUAL),
        fields=[
            FieldCreate(key="email", label="Email", data_type=FieldDataType.TEXT, is_required=True),
        ],
    )
    row = await contacts.create({"email": "[e-mail]"})

---

FastAPI integration pattern

from contextlib import asynccontextmanager
from fastapi import FastAPI
from forktex_core.database import init_engine, close_engine
from forktex_core.cache import init as cache_init, close as cache_close
from forktex_core.log import setup_logging, TraceIDMiddleware

setup_logging(service="my-api")   # call before app creation

@asynccontextmanager
async def lifespan(app: FastAPI):
    init_engine(settings.db_url, pool_size=20)
    await cache_init(settings.redis_url)
    yield
    await close_engine()
    await cache_close()

app = FastAPI(lifespan=lifespan)
app.add_middleware(TraceIDMiddleware)

---

Managed Postgres (no CREATE SCHEMA)

Library schemas (forktex_flow, forktex_grid) are isolated from your alembic by default. If your Postgres host doesn't allow CREATE SCHEMA, route them to public:

init_engine(url, schema_translate_map={
    "forktex_flow": None,    # None = public schema
    "forktex_grid": None,
})

---

Gotchas

A short list of mistakes that will save you a debugging session:

Rule	Why
Qdrant collection names: use -- not :	Qdrant rejects : with 422
Qdrant point IDs: int or str(uuid.uuid4()) only	Arbitrary strings → 400
schema_translate_map: use None key for default-schema tables	"public" key doesn't remap schema=None tables
AuditMixin requires BaseDBModel	Raises TypeError on class definition
cache.init() raises on failure	Doesn't silently degrade — handle at startup
Grid.query().fetch() returns PageResponse[GridRow]	Iterate .data, not the page itself
queue.make_worker() returns arq.Worker	Not arq.worker.WorkerSettings

Story tracks

End-to-end consumer journeys exercised against real testcontainers (no mocks). Each is the canonical answer to "what does this substrate actually do?" — read the test file, then run it locally with pytest <path>.

Story	What it proves	Path
Knowledge ingestion lifecycle	Declare a Space, upload a doc to MinIO, chunk + embed into Qdrant, walk the cross-Grid edge from a search hit back to the parent doc, archive cascades blob + vector cleanup.	tests/test_stories/test_knowledge_ingestion.py
Multi-tenant isolation	Two tenants share Postgres + Qdrant + MinIO; verified isolation across Grid.query(), namespace-prefixed Qdrant collections, and Space.to_graph() snapshots; archive in tenant A doesn't touch B.	tests/test_stories/test_multitenant_isolation.py
VECTOR storage modes	Round-trips the four storage_mode settings (none / inline / remote / both) and asserts cell shape + Qdrant state per mode.	tests/test_stories/test_vector_storage_modes.py

Each story is decomposed into act-based test methods (test_act1_declare_space → test_act5_archive_cascade) so a regression localises to its act instead of stopping the whole flow.

Reading paths

Curated entry points whether you're an LLM or a human reading cold:

• Per-module reference — one docs/<extra>.md per shipped extra (linked from each badge in the level tables above). Each page covers purpose, install, public API, quick example, and cross-refs. Module reference: database · cache · flow · vault · storage · queue · vector · grid · graph · space · api · worker · log · error · types.
• Catalog (source of truth) — src/forktex_core/catalog/catalog.json. Loaded at runtime as forktex_core.catalog.current; every README table is rendered from it.
• Examples — five runnable scripts under examples/ (graph_in_memory.py, api_minimal.py, worker_minimal.py, grid_crud.py, space_bundle.py). Each exposes a run() callable; integration tests in tests/test_examples/ keep them current.
• Story tracks — see the table above. Real testcontainers, no mocks.