Proof for every answer
your RAG ever gave.

RAG doesn't fail in the model.
It fails in the audit.

RAG compliance is not a solved problem. Retrieval works, generation works, but the moment a compliance team asks for proof of what the model saw and proof that no one tampered with it since, most RAG stacks have nothing to hand over.

In regulated industries (finance, healthcare, legal, insurance) the question is never "is your retrieval good?" It's "can you prove, for any given answer on any given day, which document was cited, what the model saw, and that no one tampered with it since?" Most RAG stacks can't.

One row per chain.
Signed, stored, surfaced.

Left: what your chain normally leaves behind. Right: what ragcompliance writes to your warehouse. No wrappers, no forks, just a callback handler passed through LangChain's config={"callbacks":[handler]} or LlamaIndex's CallbackManager.

Three lines. No chain rewrites.

The handler attaches via the standard callback channel on both frameworks. Your retriever, vector store, prompt, and LLM don't change. One callback, one row.

from langchain_openai import ChatOpenAI
from langchain_core.prompts import ChatPromptTemplate
from langchain_core.runnables import RunnablePassthrough, RunnableLambda
from ragcompliance import RAGComplianceHandler, RAGComplianceConfig

retriever = vectorstore.as_retriever(search_kwargs={"k": 4})
prompt    = ChatPromptTemplate.from_template("Context:\n{context}\n\nQ: {query}")
llm       = ChatOpenAI(model="gpt-4o-mini")

chain = (
    {"context": retriever | RunnableLambda(lambda d: "\n\n".join(c.page_content for c in d)),
     "query": RunnablePassthrough()}
    | prompt | llm
)

# the only two new lines
handler = RAGComplianceHandler(config=RAGComplianceConfig.from_env(), session_id="user-abc")
answer  = chain.invoke("Does §4.2 cover indemnification?", config={"callbacks": [handler]})

from llama_index.core import Settings
from llama_index.core.callbacks import CallbackManager
from ragcompliance import RAGComplianceConfig
from ragcompliance.llamaindex_handler import LlamaIndexRAGComplianceHandler

handler = LlamaIndexRAGComplianceHandler(
    config=RAGComplianceConfig.from_env(),
    session_id="user-abc",
)
Settings.callback_manager = CallbackManager([handler])

# any query engine now runs under the audit handler
response = query_engine.query("Does §4.2 cover indemnification?")

What ships today.
All production-hardened. All tested.

Every feature below exists because a compliance team, an SRE on-call, or a power user asked for it. No vapor. 152 tests green, including a dedicated batch and concurrency suite and a retriever-chunk regression suite for langchain-core >= 1.3.0.

Self-host. Free forever.

ragcompliance is MIT-licensed middleware. Clone it, pip install it, point it at your Supabase, and every feature on this page is yours. No paid tier, no call-home, no telemetry, no per-seat fees.

If you'd rather not run it yourself, I offer a few kinds of paid help around the project: integration reviews, SOC 2 evidence prep, custom features on contract, and an operated dashboard under your own domain. These are engagements with the author, not a locked tier of the codebase.

A tight scope, shipped in public.

RAGCompliance is a focused audit trail library for retrieval-augmented generation. Not a tracing platform, not an eval suite, not a vector database. One job: turn every RAG chain invocation into one signed, row-level-secured audit record your compliance team can actually audit.

Every checkmark below has a commit, a test suite, and a Render deploy behind it. Anything planned is marked as such, with no pretending.

Questions you'd ask before adopting this.

The first things compliance, security, and platform teams ask when they look at ragcompliance for real. Plain answers, no hedging.