RAG Pipeline

RAGPipeline is the full orchestrator. The RAG facade wraps it for the happy path.

Using the RAG facade

from synapsekit import RAG

rag = RAG(model="gpt-4o-mini", api_key="sk-...")
rag.add("Document text...")

# Async
answer = await rag.ask("Your question?")

# Streaming
async for token in rag.stream("Your question?"):
    print(token, end="")

# Sync
answer = rag.ask_sync("Your question?")

Loading documents

from synapsekit import RAG, PDFLoader, DirectoryLoader

rag = RAG(model="gpt-4o-mini", api_key="sk-...")

# From a loader (List[Document])
rag.add_documents(PDFLoader("report.pdf").load())

# Entire directory
rag.add_documents(DirectoryLoader("./docs/").load())

# Async versions
await rag.add_async("raw text string")
await rag.add_documents_async(docs)

Persistence

Save and load your vector store across sessions:

rag.save("my_store.npz")

# Later
rag.load("my_store.npz")

Using RAGPipeline directly

For full control over every component:

from synapsekit.rag.pipeline import RAGConfig, RAGPipeline
from synapsekit.llm.openai import OpenAILLM
from synapsekit.llm.base import LLMConfig
from synapsekit import SynapsekitEmbeddings, InMemoryVectorStore, Retriever
from synapsekit.memory.conversation import ConversationMemory
from synapsekit.observability.tracer import TokenTracer

llm = OpenAILLM(LLMConfig(model="gpt-4o-mini", api_key="sk-..."))
embeddings = SynapsekitEmbeddings()
store = InMemoryVectorStore(embeddings)
retriever = Retriever(store, rerank=True)

pipeline = RAGPipeline(RAGConfig(
    llm=llm,
    retriever=retriever,
    memory=ConversationMemory(window=10),
    tracer=TokenTracer(model="gpt-4o-mini"),
    retrieval_top_k=5,
    chunk_size=512,
    chunk_overlap=50,
))

await pipeline.add("Your document text...")
await pipeline.add_documents(docs)   # List[Document]

answer = await pipeline.ask("Your question?")

RAGConfig reference

Field	Type	Default	Description
llm	BaseLLM	required	LLM provider
retriever	Retriever	required	Vector store + retrieval logic
memory	ConversationMemory	required	Conversation history
tracer	TokenTracer | None	None	Token/latency tracking
retrieval_top_k	int	5	Chunks to retrieve per query
system_prompt	str	"Answer using only..."	LLM system instruction
chunk_size	int	512	Max characters per chunk
chunk_overlap	int	50	Overlap between chunks