fabra-ai 2.0.4

Context infrastructure for AI applications

Fabra

Context Infrastructure for AI Applications

Try in Browser · Quickstart · Docs

---

Fabra is the system of record for what your AI knows. We ingest, index, track freshness, and serve context data — not just query it.

This "write path ownership" enables:

• Replay any AI decision — What exactly did the model know?
• Full lineage tracking — Which features, documents, and retrievers were used?
• Freshness guarantees — Was the data stale when the decision was made?

pip install "fabra-ai[ui]"

---

Choose Your Path

ML Engineers "Feast needs Kubernetes. I just need features." from fabra import FeatureStore, entity, feature store = FeatureStore() @entity(store) class User: user_id: str @feature(entity=User, refresh="hourly") def purchase_count(user_id: str) -> int: return db.query("SELECT COUNT(*) FROM purchases WHERE user_id = ?", user_id) fabra serve features.py curl localhost:8000/features/purchase_count?user_id=123 No Kubernetes. No Spark. No YAML. Just Python. Feature Store Without K8s → · Feast vs Fabra →

AI Engineers "Someone asked what the AI knew. I couldn't tell them." from fabra import FeatureStore, context, ContextItem from fabra.retrieval import retriever store = FeatureStore() @retriever(index="docs", top_k=5) async def search_docs(query: str): pass # Auto-wired to pgvector @context(store, max_tokens=4000) async def build_prompt(user_id: str, query: str): docs = await search_docs(query) return [ContextItem(content=str(docs), priority=0)] ctx = await build_prompt("user_123", "query") print(ctx.id) # Replay this exact context anytime print(ctx.lineage) # What data was used? Context Traceability → · Compliance Guide →

---

Why Engineers Choose Fabra

1. We Own the Write Path

LangChain, Pinecone, and other tools are read-only wrappers — they query your data but don't manage it. When compliance asks "what did the AI know?", they have no answer.

Fabra ingests, indexes, and serves context data. Every decision traces back through the data that informed it.

# Replay any historical context
ctx = await store.get_context_at("01912345-6789-7abc-def0-123456789abc")
print(ctx.content)   # Exact prompt from that moment
print(ctx.lineage)   # Complete data provenance

2. Local-First, Production-Ready

Same code runs everywhere. DuckDB locally, Postgres + Redis in production.

# Development (zero setup)
fabra serve features.py

# Production (just add env vars)
FABRA_ENV=production \
FABRA_POSTGRES_URL=postgresql+asyncpg://... \
FABRA_REDIS_URL=redis://... \
fabra serve features.py

No Docker for local dev. No Kubernetes for production. Deploy to Fly.io, Railway, Cloud Run, or any container platform with one command.

3. Point-in-Time Correctness

Training ML models? We use ASOF JOIN (DuckDB) and LATERAL JOIN (Postgres) to ensure your training data reflects the world exactly as it was — no data leakage, ever.

4. Token Budget Management

No more "context too long" errors. Priority-based truncation keeps your prompts under budget.

@context(store, max_tokens=4000)
async def build_prompt(user_id: str, query: str):
    return [
        ContextItem(content=system_prompt, priority=0, required=True),
        ContextItem(content=docs, priority=1),
        ContextItem(content=history, priority=2),  # Dropped first if over budget
    ]

---

Key Capabilities

For ML Engineers

Capability	Description
Python Decorators	@feature instead of 500 lines of YAML
DuckDB + Postgres	Local dev with embedded DB, production with Postgres
Point-in-Time Joins	ASOF/LATERAL joins for training data correctness
Hybrid Features	Mix Python logic and SQL in the same pipeline
One-Command Deploy	fabra deploy fly|cloudrun|railway|render

For AI Engineers

Capability	Description
Context Accountability	UUIDv7 IDs, full lineage, replay any decision
Vector Search	Built-in pgvector with automatic chunking
Token Budgets	max_tokens with priority-based truncation
Freshness SLAs	Fail-safe when data is stale
Export	fabra context export for debugging and compliance

Production Features

• Observability: Prometheus metrics, OpenTelemetry tracing
• Reliability: Circuit breakers, fallback chains, fabra doctor
• Security: Self-hosted, your data never leaves your infrastructure

---

Architecture

Development                         Production
┌─────────────────────┐            ┌─────────────────────────┐
│  Your Python Code   │            │   Your Python Code      │
│  (@feature, @context)│            │   (@feature, @context)  │
└──────────┬──────────┘            └───────────┬─────────────┘
           │                                   │
           ▼                                   ▼
┌─────────────────────┐            ┌─────────────────────────┐
│  DuckDB (embedded)  │            │  Postgres + pgvector    │
│  In-Memory Cache    │            │  Redis                  │
└─────────────────────┘            └─────────────────────────┘

Same code. Same decorators. Different backends.
FABRA_ENV=development → FABRA_ENV=production

---

Comparison

vs Feast (Feature Store)

	Feast	Fabra
Setup	Kubernetes + Spark	pip install
Configuration	YAML	Python decorators
Time to production	Weeks	30 seconds
RAG support	None	Built-in Context Store
Traceability	None	Full lineage

Use Feast when: You have a platform team and existing K8s/Spark infrastructure.

vs LangChain (RAG)

	LangChain	Fabra
Type	Framework (orchestration)	Infrastructure (storage + serving)
Traceability	None	Full lineage + replay
Token budgets	DIY	Built-in
Data ownership	Read-only wrapper	Write path owner

Use LangChain when: You need agent orchestration and don't need compliance.

---

Get Started

pip install "fabra-ai[ui]"

# ML Engineers: Serve features
fabra serve features.py

# AI Engineers: Index documents and serve context
fabra serve chatbot.py

Try in Browser · Quickstart Guide · Full Documentation

---

Roadmap

• v1.0: Core Feature Store (DuckDB, Postgres, Redis)
• v1.2: Context Store (pgvector, retrievers, token budgets)
• v1.3: UI, Magic Retrievers, One-Command Deploy
• v1.4: Context Accountability (lineage, replay, traceability)
• v1.5: Freshness SLAs (data freshness guarantees)
• v1.6: Drift detection, RBAC, multi-region

---

Contributing

We welcome contributions! See CONTRIBUTING.md to get started.

Fabra · Apache 2.0 · 2025