kore-stack 0.1.3

Complete cognitive middleware stack for LLMs. Memory + routing + cache + observability in one install.

kore-stack

Complete cognitive middleware stack for LLMs. One install, everything connected.

Memory + Identity + Cache + Routing + Observability + A/B Testing.

The difference with LangChain/LlamaIndex: kore's routing is based on formal proof complexity theory (Selector Complexity), not heuristics.

Install

pip install kore-stack               # core (Ollama-ready)
pip install kore-stack[openai]       # + OpenAI
pip install kore-stack[anthropic]    # + Anthropic
pip install kore-stack[all]          # everything

Quick start

from kore_stack import Mind, Bridge, OllamaProvider

mind = Mind("agent.db")
llm = OllamaProvider(model="llama3.2")
bridge = Bridge(mind=mind, llm=llm, cache_ttl=3600.0)

response = bridge.think("Help me with my proof", user="carlos")

What's new in v0.1.2

• kore-mind>=0.4.0: persistent SQLite embedding cache, float16 quantization, async embeddings (ollama_embed_async), streaming batch (stream_batch / astream_batch), dimension validation in cosine_similarity
• kore-bridge>=0.3.2: dependency bump

v0.1.1

• kore-mind>=0.3.1: optimized Ollama embeddings — connection reuse, LRU cache, batch embedding, fallback warnings
• kore-bridge>=0.3.1: dependency bump

What's in the stack

kore-mind — Persistent memory engine

from kore_stack import Mind

mind = Mind("agent.db", enable_traces=True)
mind.experience("User likes Python", source="carlos")
memories = mind.recall("Python", source="carlos")

# Scoped views per user
alice = mind.scoped("alice")
alice.experience("Prefers Rust")

kore-bridge — LLM cognitive middleware

from kore_stack import Bridge, OllamaProvider

bridge = Bridge(
    mind=mind,
    llm=OllamaProvider(),
    cache_ttl=3600.0,    # smart cache
    rate_limit=3,         # cognitive rate limiting
    rate_window=3600.0,
)

# Cache hit → no LLM call. Rate limited → respond from memory.
response = bridge.think("What is P vs NP?", user="carlos")

sc-router — Routing based on Selector Complexity

from kore_stack import (
    SCRouterProvider, ToolCatalog, Tool,
    OllamaProvider, Bridge, Mind,
)
from kore_bridge.providers import OpenAIProvider

# Define your tool catalog
catalog = ToolCatalog()
catalog.register(Tool(
    name="calculator",
    description="Arithmetic calculations",
    input_types={"expression"},
    output_types={"number"},
    capability_tags={"math", "calculate"},
))

# SC routing: simple queries → local Ollama, complex → GPT-4
router = SCRouterProvider(
    providers={
        "fast": OllamaProvider(model="llama3.2"),
        "quality": OpenAIProvider(model="gpt-4o"),
    },
    catalog=catalog,
)

bridge = Bridge(mind=Mind("agent.db"), llm=router)
bridge.think("What is 2+2?")        # SC(0) → Ollama
bridge.think("Analyze market trends, cross-reference sentiment, build prediction model")
                                      # SC(2-3) → GPT-4
print(router.last_sc_level)          # 0, 1, 2, or 3

A/B Testing

from kore_stack import Experiment, OllamaProvider, Mind

exp = Experiment(
    Mind("test.db"),
    variant_a=OllamaProvider(model="llama3.2"),
    variant_b=OllamaProvider(model="mistral"),
)

result = exp.run("Explain recursion")
print(f"A: {result.time_a_ms:.0f}ms, B: {result.time_b_ms:.0f}ms, faster: {result.faster}")

Architecture

┌─────────────┐
│   Your App  │
└──────┬──────┘
       │
┌──────▼──────┐
│ kore-bridge │  Cache → Rate Limit → SC Route → LLM
│  (Bridge)   │
└──────┬──────┘
       │
┌──────▼──────┐     ┌───────────┐
│  kore-mind  │     │ sc-router │
│   (Mind)    │     │  SC(0-3)  │
│   SQLite    │     └───────────┘
└─────────────┘

Packages

Package	PyPI	What it does
kore-mind	pip install kore-mind	Memory, identity, traces, cache storage
kore-bridge	pip install kore-bridge	LLM integration, cache logic, rate limiting, A/B
sc-router	pip install sc-router	Query routing by Selector Complexity
kore-stack	pip install kore-stack	All of the above, one install

License

MIT