Non-autoregressive System 1 decision engine. Typed choice, score and yes/no decisions over any text in a single forward pass, in 100+ languages, with a router that picks the right checkpoint per request.
Python calibrationclassificationdecision-modelhuggingfacejevmodernbertmultilingualnlppythonpytorchroutingtyped-decisionszero-shot
"""Laya System 1 decision engine: LlamaIndex Integration Quickstart.
Demonstrates:
1. Sub-35ms single-choice tool routing in LlamaIndex with confidence fallback gating.
2. Multi-choice routing across multiple RAG query engines for composite questions.
3. Direct high-performance query dispatch with LayaQueryRouter.
4. Edge/Serverless deployment against remote laya-serve HTTP instances.
"""
import os
import sys
sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
from laya.integrations.llamaindex import (
LayaMultiSelector,
LayaQueryRouter,
LayaSingleSelector,
QueryBundle,
ToolMetadata,
)
# Optional mock agent for environments without local GPU / PyTorch weights:
class _DemoAgent:
def predict(self, state, questions, **kwargs):
text = str(state).lower()
q_id = next(iter(questions.keys()))
criteria = questions[q_id].get("criteria", {})
# Route database/orders queries to sql_database or database
if any(w in text for w in ("order", "orders", "account", "database", "balance", "user")):
chosen = "choice_1" if "choice_1" in criteria else "database"
probs = {"choice_0": 0.08, "choice_1": 0.88, "choice_2": 0.04}
elif "both" in text or "compare" in text or "security" in text:
chosen = "choice_0" if "choice_0" in criteria else "docs"
probs = {"choice_0": 0.52, "choice_1": 0.44, "choice_2": 0.04}
else:
chosen = "choice_0" if "choice_0" in criteria else "docs"
probs = {"choice_0": 0.91, "choice_1": 0.06, "choice_2": 0.03}
return {
"model": "laya-demo",
"answers": {
q_id: {
"choice": chosen,
"answer_confidence": probs.get(chosen, 0.90),
"confidence": probs.get(chosen, 0.90),
"probabilities": probs,
}
},
}
_agent = None
try:
from laya import Router
r = Router()
# Test if models/dependencies are present
r.predict("test", {"test": {"type": "choice", "instructions": "test", "criteria": {"a": "a"}}})
_agent = r
except Exception:
_agent = _DemoAgent()
# =====================================================================
# 1. Sub-35ms Single-Choice RAG Selector (Replaces LLMSingleSelector)
# =====================================================================
# Evaluates query against candidate query engine tools in ~33 ms without
# burning tokens or waiting 1-2 seconds for autoregressive LLM completion.
tools = [
ToolMetadata(
name="vector_documentation",
description="Semantic search over technical user documentation and API guides.",
),
ToolMetadata(
name="sql_database",
description="Structured SQL database containing customer accounts, billing, and orders.",
),
ToolMetadata(
name="summary_index",
description="High-level summaries and quarterly reports of company financial performance.",
),
]
# Configure selector with confidence threshold fallback
selector = LayaSingleSelector(
confidence_threshold=0.80,
fallback_index=0, # Fallback to vector documentation if uncertain
agent=_agent,
)
query = "Can you show me the total number of orders placed by customer #1042 last month?"
print(f"--- 1. Single-Choice Selector ---")
print(f"Query: {query}")
result = selector.select(tools, query)
selected_tool = tools[result.selections[0].index]
print(f"Selected Tool: {selected_tool.name} (Index: {result.selections[0].index})")
print(f"Reason: {result.selections[0].reason}")
# =====================================================================
# 2. Multi-Choice Selector for Composite Questions (Replaces LLMMultiSelector)
# =====================================================================
# Identifies which subset of candidate tools are relevant for complex queries.
multi_selector = LayaMultiSelector(
probability_threshold=0.25,
max_outputs=2,
agent=_agent,
)
composite_query = (
"How does the new authentication architecture documented in the guides impact "
"our customer account database security settings?"
)
print(f"\n--- 2. Multi-Choice Selector ---")
print(f"Composite Query: {composite_query}")
multi_result = multi_selector.select(tools, composite_query)
print(f"Selected {len(multi_result.selections)} relevant tools:")
for sel in multi_result.selections:
print(f" -> Tool: {tools[sel.index].name} | {sel.reason}")
# =====================================================================
# 3. Direct High-Performance Query Dispatch with LayaQueryRouter
# =====================================================================
# Dispatches incoming queries directly to backend query engines or functions.
class DummyQueryEngine:
def __init__(self, name: str):
self.name = name
def query(self, query_str: str) -> str:
return f"[{self.name}] Synthesized response for: {query_str}"
engines = {
"docs": DummyQueryEngine("DocsEngine"),
"database": DummyQueryEngine("SQLEngine"),
"finance": DummyQueryEngine("FinanceEngine"),
}
router = LayaQueryRouter(
query_engines=engines,
descriptions={
"docs": "Documentation and setup tutorials",
"database": "User transactions and database queries",
"finance": "Quarterly earnings and financial statements",
},
confidence_threshold=0.75,
fallback_key="docs",
agent=_agent,
)
db_query = "What is the account balance for user Alice?"
routed_key = router.route(db_query)
response = router.query(db_query)
print(f"\n--- 3. LayaQueryRouter Direct Dispatch ---")
print(f"Query: {db_query}")
print(f"Routed Engine Key: -> {routed_key}")
print(f"Engine Response: {response}")
# =====================================================================
# 4. Remote HTTP Server Execution (Zero Heavy Dependencies)
# =====================================================================
# Connects to your self-hosted `laya-serve` instance without requiring PyTorch on edge clients.
remote_selector = LayaSingleSelector(
base_url="http://localhost:8080",
confidence_threshold=0.85,
fallback_index=0,
)
print(f"\n--- 4. Remote HTTP Server Deployment ---")
print("Remote selector configured with base_url='http://localhost:8080' via standard urllib.")