RAGKit

Build advanced Retrieval-Augmented Generation (RAG) systems — including multi-agent RAG — in a few lines of code.

RAGKit hides the plumbing (embedders, vector stores, retrievers, agents, LLM providers) behind clean, replaceable interfaces so you can go from a folder of documents to grounded answers immediately, and swap any component when you need to.

PyPI distribution name: ragkit-vs · import name: ragkit

from ragkit import RagKit

rag = RagKit(
    pipeline="ma_rag",
    llm="groq",
    model="llama-3.3-70b-versatile",
    corpus="docs/",
).build()

response = rag.query("What is RAG?")
print(response.answer)

Table of contents

• Features
• Installation
• Quick start
• Examples
• Supported pipelines
• Supported LLM providers
• Supported document formats
• Per-agent configuration (AgentConfig)
• Custom prompts (PromptStore)
• Grounding & evidence verification
• Configuration
• Error handling
• Contributing
• FAQ
• Roadmap
• License

Features

• Provider-first, few-line API — pick a provider, point at a corpus, ask.
• Two pipelines: traditional (single-pass) and ma_rag (planner → step-definer → retriever → extractor → QA → final-answer, multi-hop).
• Per-agent configuration — with AgentConfig each MA-RAG agent can use its own LLM, model, temperature, max_tokens, and prompt; anything unset inherits the global config.
• Multiple providers, per agent — mix groq, openai, anthropic, and google across agents, or keep one provider for all.
• Prompt system (PromptStore) — default prompts ship as editable Markdown files. Override per agent inline, from a file, or from a whole directory; export the bundled defaults with export_default_prompts(...).
• Grounding / evidence verification — a two-tier gate (fast similarity pre-filter + an LLM evidence verifier) decides before any LLM call whether the retrieved context can support an answer.
• Hallucination prevention — unrelated questions are answered with "The answer is not available in the provided documents." and no agent is run (not the planner, QA, or final).
• Partial-evidence grounded reasoning — the verifier accepts context that can be combined or reasoned over to build a grounded answer, preserving MA-RAG multi-hop while still rejecting unrelated queries.
• Everything is replaceable — LLM, embedder, vector store, retriever, chunker, corpus loader, and prompts each have an interface + registry.
• Runs offline — a MockLLM + hashing embedder let the whole stack run with no API keys and no model downloads (used by the test suite).
• Introspectable — rag.agent_settings() and rag.show_prompt() report the resolved model and effective prompt for every agent (no API key needed).
• Typed (ships py.typed) and lightweight by default.

Installation

pip install ragkit-vs

The base install is tiny. Install only the extras you need:

Extra	Enables	Example
embeddings	SentenceTransformer embeddings	pip install "ragkit-vs[embeddings]"
faiss	FAISS vector store	pip install "ragkit-vs[faiss]"
ma_rag	Multi-agent pipeline (LangGraph)	pip install "ragkit-vs[ma_rag]"
pdf	PDF ingestion (pypdf)	pip install "ragkit-vs[pdf]"
groq	Groq provider	pip install "ragkit-vs[groq]"
openai	OpenAI provider	pip install "ragkit-vs[openai]"
anthropic	Anthropic provider	pip install "ragkit-vs[anthropic]"
google	Google Gemini provider	pip install "ragkit-vs[google]"
bedrock	AWS Bedrock (boto3)	pip install "ragkit-vs[bedrock]"
cerebras	Cerebras provider	pip install "ragkit-vs[cerebras]"
all	everything (dev)	pip install "ragkit-vs[all]"

Typical setup for a Groq-powered RAG over PDFs:

pip install "ragkit-vs[embeddings,faiss,ma_rag,groq,pdf]"

Quick start

1. Put your provider key in a .env file (see .env.example):

   GROQ_API_KEY=gsk_xxxxxxxx
   

2. Ask questions over your documents:

   from ragkit import RagKit
   
   rag = RagKit(pipeline="ma_rag", llm="groq", corpus="docs/").build()
   print(rag.ask("Explain Retrieval Augmented Generation"))
   

No key? Run fully offline:

from ragkit import RagKit, RagKitConfig
from ragkit.llms.mock import MockLLM

cfg = RagKitConfig()
cfg.embedding.backend = "hashing"   # deterministic, no downloads

rag = RagKit(pipeline="traditional", corpus="docs/", config=cfg, llm=MockLLM()).build()
print(rag.ask("What is in my documents?"))

Examples

Runnable scripts live in examples/:

File	What it shows
basic.py	Offline end-to-end (no keys)
traditional.py	Single-pass pipeline
ma_rag.py	Multi-agent pipeline
folder.py	Ingest a folder of mixed documents
pdf.py	Ingest a PDF
groq.py / openai.py	Choosing a provider
save_load.py	Persist and reload an index

Supported pipelines

• traditional — retrieve top-k chunks, build context, one LLM call. Fast and cheap.
• ma_rag — decomposes the question into steps, retrieves and reasons per step, then synthesizes a final answer. Better for multi-hop questions.

List them at runtime: ragkit.available_pipelines().

Supported LLM providers

groq, openai, anthropic, google (Gemini), plus a provider_manager multi-provider fallback and an offline mock. Keys are read from standard environment variables (GROQ_API_KEY, OPENAI_API_KEY, GOOGLE_API_KEY, ANTHROPIC_API_KEY) or passed via api_key=....

Supported document formats

Folders of mixed files, plus individual files: .txt, .md, .pdf, .json, .jsonl, or an in-memory list of {"title", "text"} dicts. Long documents are automatically chunked.

Per-agent configuration (AgentConfig)

Beginners write nothing extra — a single llm/model applies to every agent. Advanced users configure each MA-RAG agent independently with AgentConfig (dicts work too, for backward compatibility). Any field left unset inherits the global configuration.

from ragkit import RagKit, AgentConfig

rag = RagKit(
    pipeline="ma_rag",
    llm="groq",
    model="llama-3.3-70b-versatile",          # global default for every agent
    planner=AgentConfig(model="llama-3.3-70b-versatile", temperature=0.1),
    step_definer=AgentConfig(model="llama-3.1-8b-instant"),
    qa={"model": "llama-3.1-8b-instant"},      # dict form is also accepted
    final=AgentConfig(model="llama-3.3-70b-versatile", max_tokens=512),
    corpus="docs/",
).build()

# See exactly what each agent resolved to — no API call, no keys needed.
print(rag.agent_settings())
# {'planner': {'llm': 'groq', 'model': 'llama-3.3-70b-versatile', ...}, ...}

Each agent (planner, step_definer, extractor, qa, final) accepts llm, model, temperature, max_tokens, api_key, prompt, and prompt_file. You can even mix providers across agents (install each provider's extra and set its key):

rag = RagKit(
    pipeline="ma_rag",
    llm="groq", model="llama-3.3-70b-versatile",   # default for unset agents
    planner=AgentConfig(llm="openai", model="gpt-4o-mini"),
    qa=AgentConfig(llm="google", model="gemini-1.5-flash"),
    final=AgentConfig(llm="anthropic", model="claude-3-5-sonnet-latest"),
    corpus="docs/",
).build()

Custom prompts (PromptStore)

Default prompts ship with the package as Markdown files and are loaded automatically — you never need to specify them. To customize, override a prompt three ways (precedence: inline → file → directory → bundled default):

from ragkit import RagKit, AgentConfig

# 1) inline string
RagKit(pipeline="ma_rag", llm="groq",
       planner=AgentConfig(prompt="You are the planner...\n{question}"))

# 2) a prompt file
RagKit(pipeline="ma_rag", llm="groq",
       qa=AgentConfig(prompt_file="./prompts/qa.md"))

# 3) a whole directory — missing files fall back to the bundled defaults
RagKit(pipeline="ma_rag", llm="groq", prompt_dir="./prompts")

# traditional pipeline takes a top-level prompt / prompt_file
RagKit(pipeline="traditional", llm="groq",
       prompt="Answer only from the context.\n\n{context}\n\n{question}")

Export the bundled defaults to edit them, then load them back with prompt_dir:

import ragkit
ragkit.export_default_prompts("./prompts")   # writes planner.md, qa.md, ...

# Inspect the effective prompt an agent will actually use:
rag = RagKit(pipeline="ma_rag", llm="groq", prompt_dir="./prompts")
print(rag.show_prompt("planner"))

Custom prompts are validated on load: each agent supports a fixed set of {placeholders} (e.g. planner → {question}, qa → {goal}/{history}/ {evidence}), and an unknown placeholder raises ConfigurationError.

Grounding & evidence verification

Similarity alone measures topical relatedness, not whether the context can actually answer the question. RAGKit runs a two-tier evidence gate before any LLM call:

1. Similarity pre-filter (fast, offline) — rejects clearly-unrelated retrievals for free.
2. LLM evidence verifier — asks "can a grounded answer be constructed solely from the retrieved context?" (allowing multi-passage reasoning).

If the evidence is insufficient, RAGKit returns the standard "not available" response and runs no agents.

rag = RagKit(pipeline="ma_rag", llm="groq", corpus="docs/").build()

resp = rag.query("Who won the 2022 World Cup?")   # not in the documents
print(resp.answer)     # -> The answer is not available in the provided documents.
print(resp.evidence)   # {'accepted': False, 'decided_by': 'llm',
                       #  'verifier_verdict': 'NO', 'reason': '...'}

The decision is exposed on every response as response.evidence. Grounding is on by default and fully configurable (thresholds are calibrated for the default normalized-cosine embeddings):

RagKit(pipeline="ma_rag", llm="groq", evidence=False)                  # disable
RagKit(pipeline="ma_rag", llm="groq", evidence={"mode": "similarity"}) # offline gate only
RagKit(pipeline="ma_rag", llm="groq",
       evidence={"verifier": {"model": "llama-3.1-8b-instant"}})       # cheap verifier model

Configuration

Everything is tunable through RagKitConfig (all fields have sensible defaults):

from ragkit import RagKit, RagKitConfig

cfg = RagKitConfig()
cfg.retrieval.top_k = 8
cfg.retrieval.min_score = 0.3     # ignore weak retrievals
cfg.chunking.chunk_size = 300     # words per chunk (0 = whole document)

rag = RagKit(pipeline="ma_rag", llm="groq", corpus="docs/", config=cfg).build()

Persist and reload an index:

rag.save_index(".ragkit/my_index")
rag2 = RagKit(pipeline="traditional", llm="groq").load_index(".ragkit/my_index")

Error handling

RAGKit raises clear, typed exceptions (all subclass ragkit.RagKitError):

from ragkit import RagKit
from ragkit.exceptions import MissingAPIKeyError, CorpusNotFoundError

try:
    RagKit(pipeline="ma_rag", llm="groq", corpus="docs/").build()
except MissingAPIKeyError as e:
    print("Set your API key:", e)
except CorpusNotFoundError as e:
    print("Bad corpus path:", e)

For backward compatibility these also subclass the relevant built-ins (MissingAPIKeyError is a ValueError, CorpusNotFoundError is a FileNotFoundError, NotIndexedError is a RuntimeError, etc.).

Contributing

See CONTRIBUTING.md. In short:

pip install -e ".[all]"
pytest
ruff check .

The test suite runs offline (mock LLM + hashing embedder).

FAQ

Do I need API keys to try it? No — use the mock LLM + hashing embedder (see examples/basic.py).

Why ragkit-vs on PyPI but import ragkit? The distribution name is ragkit-vs (the ragkit name was taken); the import name stays ragkit.

Does it work on Colab / Linux / macOS / Windows? Yes — all backends ship prebuilt wheels for the major platforms.

How do I avoid installing torch? Only the embeddings extra pulls torch. Use the hashing embedder for a torch-free setup.

Roadmap

• Hybrid (dense + sparse) retrieval and reranking
• Automatic, persistent index caching
• Additional vector store backends
• CLI and optional API server

License

MIT © K Varshit