meridian-oss 1.5.2

Heroku for ML Features

Meridian

The Context Store for LLMs & ML Features

Unify RAG pipelines and ML features in a single Python decorator.

From notebook prototype to production in 30 seconds. No Kubernetes. No Spark. No YAML.

🎮 Try in Browser | 📚 Docs | 🤖 Context Store

---

The Problem

You're building an AI app. You need:

• Structured features (user tier, purchase history) for personalization
• Unstructured context (relevant docs, chat history) for your LLM
• Vector search for semantic retrieval
• Token budgets to fit your context window

Today, this means stitching together LangChain, Pinecone, a feature store, Redis, and prayer.

Meridian unifies all of this in one Python decorator.

---

The 30-Second Quickstart

pip install "meridian-oss[ui]"

from meridian.core import FeatureStore, entity, feature
from meridian.context import context, ContextItem
from meridian.retrieval import retriever

store = FeatureStore()

@entity(store)
class User:
    user_id: str

@feature(entity=User, refresh="daily")
def user_tier(user_id: str) -> str:
    return "premium" if hash(user_id) % 2 == 0 else "free"

@retriever(index="docs", top_k=3)
async def find_docs(query: str):
    pass  # Automatic vector search via pgvector

@context(store, max_tokens=4000)
async def build_prompt(user_id: str, query: str):
    tier = await store.get_feature("user_tier", user_id)
    docs = await find_docs(query)
    return [
        ContextItem(content=f"User is {tier}.", priority=0),
        ContextItem(content=str(docs), priority=1),
    ]

meridian serve features.py
# Server running on http://localhost:8000

That's it. No infrastructure. No config files. Just Python.

Context Accountability (v1.4+): Every context assembly is tracked with full lineage:

ctx = await build_prompt("user_123", "How do I upgrade?")
print(ctx.id)       # UUIDv7 identifier for replay
print(ctx.lineage)  # Full audit trail: features, retrievers, freshness

Freshness SLAs (v1.5+): Ensure your AI decisions are based on fresh data:

@context(store, max_tokens=4000, freshness_sla="5m")  # Features must be <5m old
async def build_prompt(user_id: str, query: str):
    tier = await store.get_feature("user_tier", user_id)
    return [ContextItem(content=f"User is {tier}.", priority=0)]

ctx = await build_prompt("user_123", "query")
print(ctx.is_fresh)                    # True if all features within SLA
print(ctx.meta["freshness_violations"])  # Details on any stale features

---

Why Meridian?

	Traditional Stack	Meridian
Config	500 lines of YAML	Python decorators
Infrastructure	Kubernetes + Spark + Pinecone	Your laptop (DuckDB)
RAG Pipeline	LangChain spaghetti	@retriever + @context
Feature Serving	Separate feature store	Same @feature decorator
Time to Production	Weeks	30 seconds

What Makes Meridian Different

1. One Decorator for Everything

Other tools make you choose: LangChain for RAG, Feast for features, separate vector DB. Meridian gives you @feature, @retriever, and @context — all wired together, all in Python.

2. Local-First, Production-Ready

# Development (default): DuckDB + In-Memory
MERIDIAN_ENV=development

# Production: Postgres + Redis + pgvector
MERIDIAN_ENV=production

Same code. Zero changes. Just flip an environment variable.

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

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

Automatically assembles context that fits your LLM's window. Priority-based truncation. No more "context too long" errors.

---

Key Capabilities

For AI Engineers

• Vector Search: Built-in pgvector with automatic chunking and embedding
• Magic Retrievers: @retriever auto-wires to your vector index
• Context Assembly: Token budgets, priority truncation, explainability API
• Semantic Cache: Cache expensive LLM calls and retrieval results
• Context Accountability: Full lineage tracking, context replay, and audit trails for AI decisions
• Freshness SLAs: Ensure data freshness with configurable SLAs and degraded mode handling

For ML Engineers

• Hybrid Features: Mix Python logic and SQL in the same pipeline
• Event-Driven: Trigger updates via Redis Streams
• Observability: Prometheus metrics, OpenTelemetry tracing
• Self-Healing: Circuit breakers, fallback chains, meridian doctor

For Everyone

• One-Command Deploy: meridian deploy fly|cloudrun|ecs|railway|render
• Visual UI: Dependency graphs, live metrics, context debugging
• Shell Completion: meridian --install-completion

---

Architecture

Meridian scales from laptop to production without code changes.

graph TD
    subgraph Dev [Development]
        A[Your Code] -->|Uses| B(DuckDB)
        A -->|Uses| C(In-Memory Cache)
    end

    subgraph Prod [Production]
        D[Your Code] -->|Async| E[(Postgres + pgvector)]
        D -->|Async| F[(Redis)]
    end

    Switch{MERIDIAN_ENV} -->|development| Dev
    Switch -->|production| Prod

---

Production in 60 Seconds

# Set environment variables
export MERIDIAN_ENV=production
export MERIDIAN_POSTGRES_URL=postgresql+asyncpg://...
export MERIDIAN_REDIS_URL=redis://...

# Deploy
meridian deploy fly --name my-app

Full Deployment Guide →

---

Roadmap

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

---

Get Started

pip install "meridian-oss[ui]"

Try in Browser · Quickstart Guide · Full Documentation

---

Contributing

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

Meridian · Apache 2.0 · 2025