🌟 Why StrataRAG

Modern AI applications need three things existing frameworks bolt on as afterthoughts: retrieval in the right shape (there is no one-size RAG), memory that persists and learns, and orchestration across agents. StrataRAG makes all three first-class — in a core that runs on the Python standard library alone, so your tests and CI never need a network, an API key, or a GPU.

• ✅ Five primitives — Knowledge, Memory, Tool, Agent, Pipeline. Everything composes from them.
• ✅ All ten classic RAG architectures as one-line recipes — and every recipe is an open Pipeline you can rearrange or subclass.
• ✅ Typed memory — semantic, episodic, procedural, prospective, working — read/written automatically on every agent turn.
• ✅ Multimodal GraphRAG — tables, equations, code, and images parsed as typed chunks; entity graph links evidence across modalities.
• ✅ Multi-agent orchestration — sequential workflows, hub-and-spoke routing, collaborative teams with critique rounds.
• ✅ 10 vector stores · 6 embedding providers behind two interfaces — migrate with a string change.
• ✅ Production honesty — confidence gating, per-stage tracing, eval harness, incremental ingestion, actionable errors.

🚀 Quick Start

pip install stratarag          # core: nothing else needed
pip install stratarag[all]     # every optional backend

import stratarag as sr

kb = sr.Knowledge.from_docs("docs/", chunking="markdown", graph=True)

agent = sr.Agent(
    model="claude-sonnet-4-6",            # or "echo" for offline dev
    knowledge=kb,
    memory=sr.Memory(semantic=True, episodic=True, backend="sqlite:./mem.db"),
    confidence_threshold=0.35,            # ungrounded answers get gated
)

result = agent.run("What changed in the refund policy?", user_id="u42")
print(result.output, result.confidence, result.sources)

Run python examples/05_playground_ui.py → http://localhost:7327 for a zero-dependency local playground: chat, recalled memories, sources, confidence gauge, and the stage-by-stage trace.

🏗️ The Ten RAG Architectures

Every pattern from the canonical taxonomy, each a one-liner returning an open Pipeline:

Or compose your own from the stage library — QueryRewrite, HybridRetrieve, GraphRetrieve, CorrectiveRetrieve, MultiHopRetrieve, IterativeRetrieve, Rerank, Compress, ContextFilter, MemoryRead, Generate, SelfRAGGenerate, ConfidenceGate — every stage is a plain class with run(ctx) -> ctx.

Metadata filtering works everywhere: kb.search(q, where={"source": "policy.md", "year": 2026}), per-stage defaults, or per-run overrides (pipe.run(q, where={...})). List values mean any of.

🧠 Memory Types

memory = sr.Memory(
    semantic=True,     # durable facts — "user prefers metric units"
    episodic=True,     # past runs & outcomes — learn from failures
    procedural=True,   # registered, reusable skills
    prospective=True,  # future intents that fire on time or keyword triggers
    working=True,      # rolling conversation buffer with word budget
    backend="sqlite:./mem.db",                     # or any VectorStore
    extractor="llm", model="claude-sonnet-4-6",    # LLM fact extraction
)

agent.run() calls memory.read() before answering and memory.write_turn() after. Knowledge and Memory never share a store — user context cannot pollute your source of truth.

🖼️ Multimodal GraphRAG

chunking="modality" parses tables (kept whole + linearized row-by-row), LaTeX equations, fenced code, and images (alt text + optional VLM captioner= hook) into typed chunks. graph=True builds an entity graph across all of them, so a table row and a paragraph about the same entity are graph-connected. Ingestion is incremental — re-adding a document skips unchanged chunks.

🤝 Multi-Agent Orchestration

The three enterprise deployment archetypes, with agents, pipelines, or plain callables as units:

from stratarag.orchestration import Workflow, Orchestrator, Team

# Sequential — deterministic chains (AP auditing, tax filing, underwriting)
Workflow([("ingest", extractor), ("reconcile", agent), ("comply", checker)]).run(task)

# Hub-and-spoke — a router dispatches to specialists (onboarding, maintenance)
Orchestrator({"billing": ("refunds invoices", billing_agent),
              "it": ("laptops access", it_agent)}, router=model).run(task)

# Collaborative — contribute, optionally critique each other, synthesize
Team({"siem": siem_agent, "forensics": forensics_agent},
     synthesizer=model, critique=True).run(task)

Every run returns an OrchestrationResult with a full step-by-step trace and shared state.

🗄️ Backend Matrix

One VectorStore interface, one Embedder interface, one LLMProvider interface. Specs are strings: store="qdrant:http://localhost:6333", embedder="openai:text-embedding-3-small". Missing optional packages raise MissingDependencyError with the exact pip install stratarag[extra] to run.

📊 Evals Before You Ship

report = sr.EvalSuite([
    sr.EvalCase("refund window?", expected_contains=["14 days"]),
]).run(agent)                       # Agent, Pipeline, or any callable
print(report.to_markdown())         # pass rate, faithfulness, relevance, latency, gating

🎨 Playground UI

A zero-dependency local dashboard (stratarag.dashboard.serve(agent)): chat panel, recall strip showing what the agent remembered, source cards, a confidence gauge with gated-answer styling, teach-it-a-fact input, and per-stage timing trace.

🧭 Design Principles

1. Tiny primitive set — five nouns; the ten architectures are arrangements, not new machinery.
2. Your code stays normal Python — tools are functions, stages are classes, debugging is print().
3. Layered with escape hatches — recipe → rearranged stages → subclassed stage. Moving down never requires a rewrite.
4. Offline-first — the echo model, hashing embedder, and local stores mean CI needs no network and no keys.
5. Honest failures — every boundary fails loudly with the fix in the message.

🧪 Development & Testing

python -m unittest discover -s tests    # 135 tests, no network, < 1s

The suite covers chunking edge cases, store contracts, all ten architecture recipes (behavioral assertions, not just smoke), memory types, tool failures, gating, streaming, async, orchestration archetypes, multimodal parsing, graph traversal, missing-dependency paths — plus regression tests for every bug found by dogfooding.

📄 License

Apache-2.0. See CHANGELOG.md for version history.