syrag 0.2.1

Production-oriented Python framework for building RAG services on top of FastAPI.

SyRAG

Build typed, production-oriented RAG services with a FastAPI-style developer experience.

SyRAG is a Python framework for building Retrieval-Augmented Generation services with explicit pipeline contracts, typed request/response models, and production-facing defaults on top of FastAPI.

It is designed for teams that want RAG APIs with clean provider boundaries, OpenAPI docs, request context, observability, guardrails, and testable local development without committing their whole application to a single orchestration framework.

Highlights

• FastAPI-style @app.ingest(...) and @app.query(...) route decorators.
• Typed schemas for ingest, retrieval, generation, citations, usage, and errors.
• Protocol-first extension points for chunkers, embedders, vector stores, retrieval strategies, prompt assembly, generation policies, LLMs, hooks, rate limiters, and safety guards.
• Optional integrations for OpenAI, Google, Chroma, FAISS, LangChain, and LlamaIndex.
• Development/test providers and testing helpers for reliable local and CI workflows.
• OpenAPI output, structured logging, and OpenTelemetry-compatible tracing.

Supported Python versions: 3.12+.

Installation

Core package:

pip install syrag

Optional integrations:

pip install "syrag[chroma]"
pip install "syrag[faiss]"
pip install "syrag[google]"
pip install "syrag[langchain]"
pip install "syrag[llamaindex]"
pip install "syrag[openai]"
pip install "syrag[testing]"
pip install "syrag[server]"
pip install "syrag[all]"

Quick Start

Install the runtime integrations used in this example:

pip install "syrag[chroma,openai,server]"

Set OPENAI_API_KEY in your environment before starting the app.

Create main.py:

import os
from pathlib import Path

from syrag import (
    SyRAG,
    ChromaVectorStore,
    IngestRequest,
    OpenAIEmbedder,
    OpenAILLM,
    QueryRequest,
    Settings,
)

app = SyRAG(
    title="Support Bot",
    version="0.2.1",
    description="Internal support assistant",
    settings=Settings(),
)

SUPPORT_COLLECTION = "support"

embedder = OpenAIEmbedder(
    api_key=os.environ["OPENAI_API_KEY"],
    model="text-embedding-3-small",
)
vector_store = ChromaVectorStore(
    path=Path(".syrag/chroma"),
    collection_name="support_docs",
)
llm = OpenAILLM(api_key=os.environ["OPENAI_API_KEY"], model="gpt-4.1-mini")


@app.ingest("/ingest", embedder=embedder, vector_store=vector_store)
async def ingest(request: IngestRequest) -> IngestRequest:
    """Normalize incoming documents before they enter the ingest pipeline."""
    return request.model_copy(
        update={
            "collection": request.collection or SUPPORT_COLLECTION,
            "metadata": {"source": "api", **request.metadata},
        }
    )


@app.query("/query", embedder=embedder, vector_store=vector_store, llm=llm)
async def query(request: QueryRequest) -> QueryRequest:
    """Apply route-level retrieval defaults before generation."""
    return request.model_copy(
        update={
            "collection": request.collection or SUPPORT_COLLECTION,
            "top_k": min(request.top_k, 5),
        }
    )

Serve the app with any ASGI server. With the server extra installed:

uvicorn main:app.api --reload

The framework exposes:

• POST /ingest
• POST /query
• GET /health
• OpenAPI docs at /docs

Ingest a document:

curl -X POST http://127.0.0.1:8000/ingest \
  -H "content-type: application/json" \
  -d '{"documents":["SyRAG builds typed RAG services."],"collection":"demo"}'

Query it:

curl -X POST http://127.0.0.1:8000/query \
  -H "content-type: application/json" \
  -d '{"query":"What does SyRAG build?","collection":"demo","top_k":1}'

Extension Points

SyRAG keeps the pipeline explicit. You can swap or extend:

• Chunker
• Embedder
• VectorStore
• RetrievalStrategy
• PromptAssembler
• GenerationPolicy
• LLM
• RequestContextHook
• AuthHook
• RateLimiter
• SafetyGuard

First-Party Providers

Core:

• InMemoryEmbedder
• InMemoryVectorStore
• InMemoryLLM
• PassThroughChunker
• SQLiteVectorStore

InMemoryEmbedder, InMemoryVectorStore, and InMemoryLLM are development/test utilities. They are not intended as production embedding, retrieval, or generation backends.

Optional chroma extra:

• ChromaVectorStore

Optional faiss extra:

• FAISSVectorStore

Optional google extra:

• GoogleEmbedder
• GoogleLLM

Optional openai extra:

• OpenAIEmbedder
• OpenAILLM

Observability And Operations

SyRAG includes:

• structured error responses with stage information
• request-scoped RequestContext with request IDs and tenant IDs
• StructuredLogging and JSONLogFormatter
• OpenTelemetryTracing built on the OpenTelemetry API package
• request throttling via InMemoryRateLimiter
• payload validation via DefaultSafetyGuard

Testing

Install the testing extra to use:

• create_test_app(...)
• create_test_client(...)
• seed_documents(...)
• fake providers such as FakeEmbedder, FakeVectorStore, and FakeLLM

Docs

• Hosted docs
• Docs index
• Overview
• Architecture
• Component contracts
• Public API
• Compatibility matrix
• Cookbook
• Provider examples
• 0.2.0 roadmap
• Releasing
• MVP status

Build the docs locally with MkDocs Material:

uv run --group docs mkdocs serve

Examples

• Minimal app

Cookbook

• Qdrant vector store pipeline
• LangChain + Qdrant RAG
• LangChain agent with SyRAG tool
• LangGraph + SyRAG RAG workflow
• LlamaIndex + Qdrant RAG

Full example scripts live in examples/integrations.

Community

• Contributing guide
• Support
• Security policy
• Code of conduct
• Changelog