Skip to main content

Build a RAG assistant with LangChain and VectorAmp

Use the VectorAmp LangChain integration when you want LangChain orchestration with VectorAmp handling hosted embeddings, SABLE-backed retrieval, metadata filters, and dataset storage.

What you will build​

A small RAG assistant that retrieves context from a VectorAmp dataset and passes the context into a LangChain chain.

Prerequisites​

  • Python 3.10+.
  • A VectorAmp API key.
  • A dataset ID with documents already ingested, or documents you will add from Python.
  • An LLM integration for generation.

Step 1: Install packages​

pip install langchain-vectoramp langchain

If you use OpenAI for the generator:

pip install langchain-openai

Step 2: Configure the vector store​

from langchain_vectoramp import VectorAmpVectorStore

store = VectorAmpVectorStore(
    api_key="vsk_...",            # or set VECTORAMP_API_KEY
    dataset_id="ds_123",
)

Step 3: Add documents or use existing ingested content​

For small examples, add text directly:

store.add_texts(
    [
        "VectorAmp uses SABLE for filter-native vector search.",
        "Metadata filters are planned before retrieval, not bolted on after.",
    ],
    metadatas=[
        {"source": "architecture", "team": "platform"},
        {"source": "architecture", "team": "platform"},
    ],
)

For production, prefer VectorAmp source ingestion for Drive, S3, GCS, web, Jira, Confluence, or file uploads.

Step 4: Retrieve with filters​

docs = store.similarity_search(
    "How does VectorAmp handle metadata filters?",
    k=5,
    filter={"team": "platform"},
)

for doc in docs:
    print(doc.page_content)
    print(doc.metadata)

Step 5: Use the retriever in a chain​

from langchain_core.prompts import ChatPromptTemplate
from langchain_openai import ChatOpenAI

retriever = store.as_retriever(
    search_kwargs={"k": 5, "filter": {"team": "platform"}}
)

prompt = ChatPromptTemplate.from_messages([
    ("system", "Answer using only the provided context. Cite source metadata when possible."),
    ("human", "Question: {question}\n\nContext:\n{context}"),
])

llm = ChatOpenAI(model="gpt-4o-mini")

question = "How does VectorAmp handle metadata filters?"
context_docs = retriever.invoke(question)
context = "\n\n".join(doc.page_content for doc in context_docs)
answer = (prompt | llm).invoke({"question": question, "context": context})
print(answer.content)

Production notes​

  • Keep tenant and role filters on the server side when building user-facing apps.
  • Store source IDs, URLs, or document IDs in metadata so answers can link back to original content.
  • Use VectorAmp ingestion for large or frequently updated corpora.
  • Treat the LLM answer as synthesis; VectorAmp retrieval provides the evidence.

Validation checklist​

  • Retrieval returns expected documents before adding the LLM.
  • Filters exclude documents from other teams or tenants.
  • Generated answers include source references.
  • Failure modes are clear when no relevant context is found.