Raglan

A lightweight, highly configurable RAG retrieval engine. Framework-free, protocol-driven.

Why Raglan?

Standard RAG has one step: vector search → top-K results. Real-world retrieval needs more:

Six-Stage Pipeline

User query
  │
  ▼
┌─────────────────────────────────┐
│ Stage 1: QueryExpander           │  ← LLM generates entity extraction + 3 variants
└──────────────┬──────────────────┘
               │ original + variants = parallel search
               ▼
┌─────────────────────────────────┐
│ Stage 2: Retrievers (parallel)   │
│  · Dense: pgvector / Qdrant / ChromaDB / ...
│  · Sparse: BM25 full-text        │
└──────────────┬──────────────────┘
               │ multi-source results
               ▼
┌─────────────────────────────────┐
│ Stage 3: Fusion                  │  ← RRF / Weighted / RoundRobin
└──────────────┬──────────────────┘
               │ fused candidates
               ▼
┌─────────────────────────────────┐
│ Stage 4: Reranker (optional)     │  ← Cross-Encoder / Cohere Rerank
│ Filters scores below threshold   │
└──────────────┬──────────────────┘
               │ top-N candidates
               ▼
┌─────────────────────────────────┐
│ Stage 5: ContextBuilder          │  ← Parent expansion / window / passthrough
│ Greedy fill up to max_tokens     │
└──────────────┬──────────────────┘
               │
               ▼
         Final Top-N results

Installation

pip install raglan-retrieval

With optional providers:

pip install raglan-retrieval[openai]        # OpenAI embedder + expander
pip install raglan-retrieval[pgvector]      # Postgres + pgvector retriever
pip install raglan-retrieval[huggingface]   # HF embedder + Cross-Encoder reranker
pip install raglan-retrieval[qdrant]        # Qdrant retriever
pip install raglan-retrieval[chromadb]      # ChromaDB retriever
pip install raglan-retrieval[cohere]        # Cohere reranker
pip install raglan-retrieval[dashscope]     # Alibaba DashScope embedder
pip install raglan-retrieval[litellm]       # LiteLLM multi-provider expander
pip install raglan-retrieval[all]           # Everything

5-Minute Quickstart

from raglan import Raglan
from raglan.retrievers import BM25Retriever
import asyncio


async def main():
    # Step 1: Set up a retriever with your data
    bm25 = BM25Retriever()

    async def chunks():
        yield [
            ("doc1", "Return policy: items can be returned within 30 days.", None),
            ("doc2", "Refund process: refunds are issued to the original payment method.", None),
            ("doc3", "Shipping: orders ship within 2 business days.", None),
        ]

    await bm25.index(chunks())

    # Step 2: Build the pipeline (all stages optional beyond retrievers)
    rag = Raglan.builder().with_retrievers([bm25]).build()

    # Step 3: Search
    results, trace = await rag.search("how to return my order")

    for r in results:
        print(f"[{r.score:.3f}] {r.content}")

    print(f"Pipeline took {trace.total_ms:.1f}ms")


asyncio.run(main())

Adding vector search and reranking

from raglan.context_builders import ParentExpander
from raglan.embedders import OpenAIEmbedder
from raglan.expanders import OpenAIExpander
from raglan.rerankers import CrossEncoderReranker
from raglan.retrievers import ConfigurablePgvectorRetriever

rag = (
    Raglan.builder()
    .with_expander(OpenAIExpander(model="gpt-4o-mini"))
    .with_embedder(OpenAIEmbedder(model="text-embedding-3-small"))
    .with_retrievers(
        [
            ConfigurablePgvectorRetriever(
                connection_string="postgresql://...",
                table="kb.chunks",
                embedding_column="embedding",
                parent_id_column="parent_id",
            ),
            BM25Retriever(),
        ]
    )
    .with_reranker(
        CrossEncoderReranker(
            model_name="ms-marco-TinyBERT-L2-v2",
            min_score=0.5,
        )
    )
    .with_context_builder(
        ParentExpander(
            loader=my_parent_loader,
            max_tokens=6000,
        )
    )
    .with_fallback_mode("degrade")
    .build()
)

results, trace = await rag.search("damaged order return policy")

Design Philosophy

1. Framework-Free — No dependency on LangChain, LlamaIndex, or any specific vector database or LLM provider. Protocols define the interfaces; you provide the implementations.

2. Graceful Degradation — Every stage is independent. Query expansion fails? Use the original query. Cross-Encoder not installed? Skip reranking. BM25 unavailable? Pure vector search still works.

3. Fully Configurable — Every stage's parameters, weights, models, and thresholds are configurable via the Builder or from_dict().

4. Production-Ready — Extracted from production systems processing thousands of support emails daily. Handles extreme text, multilingual, and high-concurrency scenarios.

Comparison with Existing RAG Frameworks

Project Structure

raglan/
├── README.md
├── docs/
│   ├── architecture.md      # Architecture design
│   ├── configuration.md     # Full configuration reference
│   ├── pipeline.md          # Six-stage deep dive
│   └── examples.md          # Multi-scenario examples
├── raglan/
│   ├── __init__.py
│   ├── raglan.py            # Raglan facade + Builder
│   ├── pipeline.py          # Pipeline engine + stage dispatch
│   ├── protocols.py         # User-implementable abstract interfaces
│   ├── types.py             # ScoredChunk, SearchResult, Filter, etc.
│   ├── exceptions.py        # Exception hierarchy
│   ├── observability.py     # Metrics collector
│   ├── expanders/           # Query expansion (OpenAI, LiteLLM, Identity)
│   ├── embedders/           # Embedding (OpenAI, HuggingFace, DashScope)
│   ├── retrievers/          # Search backends (BM25, pgvector, Qdrant, ChromaDB)
│   ├── fusion/              # Result fusion (RRF, Weighted, RoundRobin)
│   ├── rerankers/           # Reranking (CrossEncoder, Cohere)
│   ├── context_builders/    # Context assembly (Parent, Window, Passthrough)
│   ├── middleware/           # Timeout, Retry, CircuitBreaker, Logging
│   └── resilience/          # RateLimiter, RetryBudget, HealthChecker
└── tests/

License

MIT