largestack 1.0.0

Largestack AI — production-grade candidate framework for typed agents, tools, RAG, guardrails, and orchestration

Largestack AI

Largestack AI is an agentic AI framework for building practical AI applications with agents, tools, workflows, RAG, guardrails, observability, and deployment support in one project.

It is designed for developers who want to build real AI systems without starting from a blank file: support-ticket agents, RAG assistants, code reviewers, workflow automations, BFSI governance flows, and enterprise-style AI copilots.

Current status: v1.0 Release Candidate / controlled-pilot ready. Ubuntu, Mac evidence, Windows clean validation, Docker, security, package, DeepSeek live validation, and 24-hour soak evidence have passed.

---

Why Largestack?

Most agent frameworks solve only one layer: agents, chains, RAG, or observability. Largestack brings the main production surfaces together:

Layer	What Largestack provides
Agents	Agent, typed agents, role-based agents, multi-agent teams
Tools	Safe tool calling, schemas, retries, timeout controls, approval policies
Workflows	Sequential, parallel, router, supervisor, graph/DAG-style orchestration
RAG	Loaders, chunking, retrievers, rerankers, vector stores, citations, no-answer behavior
Guardrails	PII checks, injection controls, topic/sensitive data policies, tool/provider policies
Memory	Buffer, long-term, vector-backed, shared and isolated memory patterns
Observability	Traces, cost tracking, event logs, dashboard APIs, OTEL helpers
Enterprise	RBAC, audit trail, tenant scoping, SSO/session modules, payment/billing scaffolds
Deployment	Docker, Compose, Helm charts, CI validation, release evidence
Testing	Unit, integration, security, RAG eval, live provider validation, generated project checks

---

5-minute quickstart

1. Get the source

# Public GitHub clone URL should be added after repository visibility is enabled.
cd largestack

2. Create environment

python3.12 -m venv .venv
source .venv/bin/activate
python -m pip install -U pip setuptools wheel

3. Install

For normal development:

python -m pip install -e ".[dev]"

For CPU-only PyTorch dependency resolution on Linux/macOS:

PIP_EXTRA_INDEX_URL=https://download.pytorch.org/whl/cpu \
python -m pip install -e ".[dev]"

4. Run a first validation

python -m pytest tests/unit/test_memory.py -q --tb=short

5. Run the full suite

python -m pytest tests -q --tb=short -ra

---

Minimal agent example

import asyncio
from largestack import Agent

async def main():
    agent = Agent(
        name="assistant",
        llm="deepseek/deepseek-chat",
        instructions="Be concise and practical."
    )
    result = await agent.run("Explain Largestack in one sentence.")
    print(result.content)

asyncio.run(main())

For deterministic tests, use the built-in test/offline model patterns instead of a live cloud provider.

---

Live provider setup

DeepSeek:

export LARGESTACK_DEEPSEEK_API_KEY="your_key_here"
python examples/01_hello/main.py

OpenAI:

export LARGESTACK_OPENAI_API_KEY="your_key_here"
export LARGESTACK_DEFAULT_MODEL="openai/gpt-4o-mini"
python examples/01_hello/main.py

Never commit .env or paste API keys into source files.

---

LLM/API provider support

Largestack is provider-switchable. The core agent, workflow, RAG, guardrail, and observability layers run through a model string such as openai/gpt-4o-mini, anthropic/claude-sonnet-4-6, deepseek/deepseek-chat, litellm/groq/llama-3.1-70b-versatile, or local/llama3.2.

Recommended public claim:

Largestack supports OpenAI/GPT, Anthropic/Claude, DeepSeek, LiteLLM, Ollama/local models, and many OpenAI-compatible providers through a verified/partial capability matrix.

Do not claim every provider has identical production-grade tool calling, structured output, streaming, and cost tracking until that provider/model has passed live E2E validation.

Provider/API path	Model string example	Env/config	Status
OpenAI / GPT	openai/gpt-4o-mini	LARGESTACK_OPENAI_API_KEY	Verified primary adapter path
Anthropic / Claude	anthropic/claude-sonnet-4-6	LARGESTACK_ANTHROPIC_API_KEY	Verified native adapter path
DeepSeek	deepseek/deepseek-chat	LARGESTACK_DEEPSEEK_API_KEY	Live E2E validated
LiteLLM gateway	litellm/<provider>/<model>	Provider-specific LiteLLM env vars	Partial; downstream capability varies
Local OpenAI-compatible	local/<model>	LARGESTACK_OPENAI_COMPATIBLE_BASE_URL	Partial; gateway/model capability varies
Ollama native	ollama/<model>	LARGESTACK_OLLAMA_BASE_URL optional	Partial; chat path first
Azure OpenAI	azure/<deployment>	LARGESTACK_AZURE_OPENAI_KEY, LARGESTACK_AZURE_OPENAI_ENDPOINT	Partial; deployment-specific
Groq, Mistral, OpenRouter, xAI, Cerebras, SambaNova, NVIDIA	<provider>/<model>	LARGESTACK_<PROVIDER>_API_KEY	Partial/OpenAI-compatible; verify live
Google/Gemini, Cohere, Bedrock	<provider>/<model>	Provider env/credentials	Partial; feature support differs

Inspect the runtime matrix:

python - <<'PY'
from largestack import provider_support_matrix
for row in provider_support_matrix():
    print(row["provider"], row["status"], "tools=", row["tool_calling"], "structured=", row["structured_output"])
PY

Run the provider-switchable flow demo offline:

python examples/provider_flow_demo/main.py

Run the same flow against GPT:

export LARGESTACK_OPENAI_API_KEY="your_key_here"
export LARGESTACK_DEFAULT_MODEL="openai/gpt-4o-mini"
export LARGESTACK_FLOW_DEMO_LIVE=1
python examples/provider_flow_demo/main.py

Run the same flow against Claude:

export LARGESTACK_ANTHROPIC_API_KEY="your_key_here"
export LARGESTACK_DEFAULT_MODEL="anthropic/claude-sonnet-4-6"
export LARGESTACK_FLOW_DEMO_LIVE=1
python examples/provider_flow_demo/main.py

Run the same flow against a local OpenAI-compatible endpoint:

export LARGESTACK_OPENAI_COMPATIBLE_BASE_URL="http://localhost:11434/v1"
export LARGESTACK_OPENAI_COMPATIBLE_API_KEY="ollama"
export LARGESTACK_DEFAULT_MODEL="local/llama3.2"
export LARGESTACK_FLOW_DEMO_LIVE=1
python examples/provider_flow_demo/main.py

---

Flow demo

The quickest workflow demo is examples/provider_flow_demo/main.py. It runs offline by default and can be switched to any configured provider by changing only LARGESTACK_DEFAULT_MODEL.

flowchart LR
    U[User task] --> I[Intake agent]
    I --> P[Planner agent]
    P --> R[Responder agent]
    R --> O[Final answer]

What the demo proves:

• one task flows through three agents,
• DAG dependencies control execution order,
• each agent can use the same model string or provider family,
• offline TestModel validation requires no API key,
• live mode works with GPT, Claude, DeepSeek, LiteLLM, or local-compatible providers when credentials are configured.

---

Built-in example areas

Example	Purpose
examples/00_offline_test_model.py	Offline deterministic model check
examples/01_hello	Basic provider-backed agent
examples/02_tools	Tool calling
examples/03_team	Multi-agent/team behavior
examples/04_guards	Guardrails/security behavior
examples/05_rag_knowledge	RAG with knowledge files
examples/06_streaming	Streaming responses
examples/07_structured	Structured outputs
examples/08_mcp_server	MCP server pattern
examples/10_full_app	Integrated app pattern
examples/provider_flow_demo	Provider-switchable workflow demo
examples/rag_basic	Basic RAG assistant
examples/fintech_kyc	BFSI/KYC style workflow
examples/riva_ai	Riva/Largestack demo pipelines

---

Validation status

Latest confirmed release-candidate evidence includes:

Gate	Status
Ubuntu full pytest	Passed
Mac validation	Passed / evidence added
Windows validation	Passed / clean Windows validation confirmed
DeepSeek live difficult projects	5/5 passed
Full DeepSeek integration suite	Passed with one known provider-format skip
Provider support matrix	Present / explicit verified-partial-adapter statuses
Offline provider flow demo	Passed with TestModel
Security suite	Passed
RAG eval suite	Passed
Package build + twine check	Passed
Docker runtime /health	Passed
Helm lint/template	Passed
4-hour soak evidence	Passed
24-hour soak	Passed / 210 successful cycles / 0 recorded failures

---

Architecture at a glance

flowchart TD
    U[User / API / CLI / App] --> C[CLI or SDK]
    C --> A[Agent Runtime]
    A --> W[Workflow Orchestrator]
    A --> T[Tool Registry]
    A --> M[Memory Layer]
    A --> R[RAG Layer]
    A --> G[Guardrails]
    W --> S[State / Checkpoints]
    T --> I[Integrations]
    R --> V[Vector Stores / Retrievers / Rerankers]
    G --> E[Enterprise Policies]
    A --> O[Observability]
    O --> D[Dashboard / Metrics / Traces]
    E --> AUD[Audit / RBAC / Tenant Controls]
    C --> DEP[Docker / Compose / Helm]

---

Repository map

Path	Purpose
largestack/_core	Main agent/tool/runtime primitives
largestack/_workflow	Workflow graph, checkpoints, interrupts, subgraphs
largestack/_rag	RAG query engines, eval, summary index
largestack/_memory	Memory stores and memory tools
largestack/_guard	Provider/tool guardrail policies
largestack/_security	Sandbox, permissions, vault, encryption, network controls
largestack/_enterprise	RBAC, audit, tenant, SSO/session, billing/payment modules
largestack/_observe	Cost, traces, OTEL, telemetry helpers
largestack/_dashboard	Dashboard app and APIs
largestack/_integrations	Provider/tool integrations
largestack/_templates	Project starter templates
examples/	Runnable examples
tests/	Unit, integration, security, RAG eval tests
scripts/	Certification, smoke, scenario, and live DeepSeek validation scripts
deploy/	Docker, Compose, Helm, monitoring assets
release_evidence/	Validation evidence and release proof

---

Production-positioning honesty

Largestack is strong for:

• developer demos,
• investor demos,
• internal AI platform experiments,
• controlled pilots,
• agentic framework portfolio proof,
• private beta deployments.

Largestack should not yet be marketed as:

• fully BFSI-certified,
• SOC2/ISO-certified,
• full LangChain/LangGraph ecosystem replacement,
• public SaaS production platform without load tests, external VAPT, and real Kubernetes install proof.

Known limitations are tracked in docs/known-limitations.md. Review that file before publishing release, SaaS, BFSI, or regulated-enterprise claims.

---

Roadmap

Priority	Work
P0	Add load/concurrency evidence after completed 24h soak
P0	Queue/backpressure for high traffic
P0	Distributed workers and job leasing
P0	Durable replay/checkpoint recovery
P1	Real Kubernetes cluster install test
P1	Observability UI polish and replay debugger
P1	More beginner templates and tutorials
P2	Public docs website
P2	Community examples and plugin ecosystem
P3	Enterprise certifications, VAPT, compliance evidence

---

License

Apache-2.0.