fabra-ai 2.1.1

Context infrastructure for AI applications

Fabra

Stop spending hours debugging AI decisions you can't reproduce.

When your AI gives a bad answer, you need to know: What did it see? What got dropped?
Can I replay this exact decision? Fabra gives you the answer.

---

The Problem

Incident risk: Your AI shipped a bad answer and you can't explain why.

Credibility risk: Support escalations are vibes and screenshots — not fixable tickets.

Velocity risk: You're scared to ship because you can't debug regressions.

Every AI team hits this wall: you can't fix what you can't reproduce.

---

30 Seconds to Proof

pip install fabra-ai && fabra demo

That's it. Server starts, makes a test request, shows you the result. No Docker. No config files. No API keys.

What you'll see

  Fabra Demo Server

  Testing feature retrieval...
  curl localhost:8000/features/user_engagement?entity_id=user_123

  Response:
  {
    "value": 87.5,
    "freshness_ms": 0,
    "served_from": "online"
  }

  Press Ctrl+C to stop, or visit http://localhost:8000/docs

---

The Solution: Context Records

Fabra creates a Context Record for every AI decision — an immutable snapshot that turns "the AI was wrong" into a fixable ticket.

ctx = await build_prompt("user_123", "how do I get a refund?")

print(ctx.id)       # ctx_018f3a2b-... (your receipt)
print(ctx.lineage)  # Exactly what data was used

That ctx.id is permanent. Days later, you can:

# See exactly what the AI knew
fabra context show ctx_018f3a2b-...

# Compare two decisions side-by-side
fabra context diff ctx_a ctx_b

This is the difference between "we think it worked" and "here's the proof."

---

Two Entry Points, One Infrastructure

For ML Engineers "I need features in production, not a platform team." from fabra import FeatureStore, entity, feature from datetime import timedelta store = FeatureStore() @entity(store) class User: user_id: str @feature(entity=User, refresh=timedelta(hours=1)) 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?entity_id=u123 # {"value": 47, "freshness_ms": 0, "served_from": "online"} Python decorators. Not YAML.

For AI Engineers "Compliance asked what the AI knew. I need an answer." from fabra import FeatureStore from fabra.context import context, ContextItem store = FeatureStore() @context(store, max_tokens=4000, freshness_sla="5m") async def build_prompt(user_id: str, query: str): tier = await store.get_feature("user_tier", user_id) docs = await search_docs(query) return [ ContextItem(content=f"User tier: {tier}", priority=1), ContextItem(content=docs, priority=2), ] ctx = await build_prompt("user_123", "question") print(ctx.id) # ctx_018f3a2b-... (stable Context Record ID) print(ctx.lineage) # exact data used, full provenance Full audit trail. Not a black box.

---

Why It Works

1. You Own Your Data

LangChain queries your vector DB. Fabra is your vector DB. We ingest, index, track freshness, and serve. When someone asks "what did the AI know?", we have the answer because we never lost sight of the data.

# Replay any historical context
fabra context show ctx_018f3a2b-7def-7abc-8901-234567890abc

# Compare what changed between two decisions
fabra context diff ctx_abc123 ctx_def456

2. Same Code Everywhere

Development uses DuckDB (zero setup). Production uses Postgres + Redis (just add env vars). Your feature definitions don't change.

# Development (right now, on your laptop)
fabra serve features.py

# Production (same code, different backends)
FABRA_ENV=production \
FABRA_POSTGRES_URL=postgresql://... \
FABRA_REDIS_URL=redis://... \
fabra serve features.py

3. Point-in-Time Correctness

Training ML models? We use ASOF JOIN to ensure your training data reflects exactly what the model would have seen at prediction time. No data leakage. No training-serving skew.

4. Token Budgets That Work

No more prompt length errors in production. Set a budget, assign priorities, and low-priority items get dropped automatically.

@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=user_history, priority=1),  # dropped first if over budget
        ContextItem(content=docs, priority=2),
    ]

---

What's Real

This isn't a framework that wraps other tools. This is infrastructure:

Capability	What It Does
Feature Store	@feature decorators, online/offline stores, point-in-time joins
Context Store	@context decorators, token budgeting, lineage tracking
Vector Search	Built-in pgvector, automatic chunking, freshness tracking
Context Replay	fabra context show <id> returns exact historical state
Context Diff	fabra context diff <id1> <id2> shows what changed
Freshness SLAs	freshness_sla="5m" fails if data is stale
Diagnostics	fabra doctor validates your setup

CLI

fabra serve features.py      # Start the server
fabra demo                   # Interactive demo (no setup)
fabra doctor                 # Diagnose configuration issues
fabra context show <id>      # Replay historical context
fabra context diff <a> <b>   # Compare two contexts
fabra context list           # List recent contexts
fabra context export <id>    # Export for audit
fabra deploy fly|railway     # Generate deployment config
fabra ui features.py         # Launch the dashboard (requires Node.js)

Note: fabra ui requires Node.js. Run npm install in src/fabra/ui-next/ if dependencies aren't installed.

---

How Fabra Fits in Your Stack

Fabra is not a replacement for:

Tool	Purpose	Relationship to Fabra
Airflow / Dagster	Batch workflow orchestration	Use for pipelines that feed Fabra
MLflow / W&B	Model training & experiment tracking	Use for training; Fabra handles inference-time context
LangChain / LlamaIndex	LLM orchestration & chains	Use for orchestration; Fabra provides the data layer

Fabra replaces or complements:

Tool	Fabra Advantage
Feast	Simpler setup, built-in context assembly
Custom feature serving	Production-ready out of the box
Ad-hoc RAG pipelines	Lineage, freshness SLAs, token budgets

See full comparison guide for detailed breakdowns vs Feast, Tecton, LangChain, and Pinecone.

---

Honest Comparison

vs Feast

	Feast	Fabra
Can you prove what happened?	Partial (features only)	Full Context Record
Can you replay a decision?	No	Yes, built-in
Setup time	Days/Weeks	30 seconds
Infrastructure	Kubernetes	None required

Choose Feast if: You have a platform team and existing K8s infrastructure.

vs LangChain

	LangChain	Fabra
Can you prove what happened?	No	Full Context Record
Can you explain why it missed something?	No	Yes, dropped items logged
Can you replay a decision?	Manual	Built-in
Data ownership	Queries external stores	Owns the write path

Choose LangChain if: You need agent chains and don't need compliance.

Note: You can use Fabra + LangChain together — Fabra for storage/serving, LangChain for orchestration.

---

Production Checklist

• Observability: Prometheus metrics at /metrics, structured logging
• Reliability: Circuit breakers, fallback chains, health checks
• Security: Self-hosted, your data stays in your infrastructure
• Deployment: One-command deploy to Fly.io, Railway, Cloud Run, Render

---

Quick Start (Detailed)

Feature Store

pip install fabra-ai
fabra serve examples/demo_features.py

Test it:

curl localhost:8000/features/user_engagement?entity_id=user_123

Response:

{"value": 87.5, "freshness_ms": 0, "served_from": "online"}

Context Store

pip install fabra-ai
fabra serve examples/demo_context.py

Test it:

curl -X POST localhost:8000/v1/context/chat_context \
  -H "Content-Type: application/json" \
  -d '{"user_id":"user_123","query":"how do features work?"}'

Response:

{
  "id": "ctx_018f3a2b-...",
  "content": "You are a helpful AI assistant...",
  "meta": {
    "freshness_status": "guaranteed",
    "token_usage": 150
  },
  "lineage": {
    "features_used": ["user_tier", "user_engagement_score"],
    "retrievers_used": ["demo_docs"]
  }
}

Replay it later:

fabra context show ctx_018f3a2b-...

---

What Fabra Does Not Do

• Agent orchestration - Use LangChain
• Workflow scheduling - Use Airflow/Dagster
• High-QPS streaming inference - Use Tecton
• No-code builders - This is Python infrastructure

Fabra focuses on one thing: turning "the AI was wrong" into a fixable ticket.

---

Try in Browser · Quickstart · Docs

---

Fabra · Apache 2.0 · 2025

Stop bleeding time and credibility when AI misbehaves.