dolev31

dolev31/lasagna

Swap one RAG pipeline stage, hold the rest fixed, and measure whether the gain survives composition. The framework behind "Retrieval Augmented Generation Compositionality": 450 configurations, eight benchmarks, corpora from 5K to 200K documents.

20 stars 2 forks 20 watchers Python Apache License 2.0
ablation-study agentic-rag beir benchmark evaluation hybrid-search information-retrieval langgraph llm mcp mcp-server python rag rag-evaluation ragas reproducibility reranking retrieval-augmented-generation semantic-search vector-database
View on GitHub Website
26 Open Issues Need Help Last updated: Oct 26, 2026

Open Issues Need Help

View All on GitHub
Wanted: your paper's component, in Lasagna about 17 hours ago
help wanted component-wanted size: feature
dolev31/lasagna
20

Swap one RAG pipeline stage, hold the rest fixed, and measure whether the gain survives composition. The framework behind "Retrieval Augmented Generation Compositionality": 450 configurations, eight benchmarks, corpora from 5K to 200K documents.

Python
#ablation-study#agentic-rag#beir#benchmark#evaluation#hybrid-search#information-retrieval#langgraph#llm#mcp#mcp-server#python#rag#rag-evaluation#ragas#reproducibility#reranking#retrieval-augmented-generation#semantic-search#vector-database
Research: does composition hold with an open-weight generator? about 17 hours ago
help wanted research size: research
dolev31/lasagna
20

Swap one RAG pipeline stage, hold the rest fixed, and measure whether the gain survives composition. The framework behind "Retrieval Augmented Generation Compositionality": 450 configurations, eight benchmarks, corpora from 5K to 200K documents.

Python
#ablation-study#agentic-rag#beir#benchmark#evaluation#hybrid-search#information-retrieval#langgraph#llm#mcp#mcp-server#python#rag#rag-evaluation#ragas#reproducibility#reranking#retrieval-augmented-generation#semantic-search#vector-database
Research: corpus scaling with out-of-domain distractors about 17 hours ago
help wanted research size: research
dolev31/lasagna
20

Swap one RAG pipeline stage, hold the rest fixed, and measure whether the gain survives composition. The framework behind "Retrieval Augmented Generation Compositionality": 450 configurations, eight benchmarks, corpora from 5K to 200K documents.

Python
#ablation-study#agentic-rag#beir#benchmark#evaluation#hybrid-search#information-retrieval#langgraph#llm#mcp#mcp-server#python#rag#rag-evaluation#ragas#reproducibility#reranking#retrieval-augmented-generation#semantic-search#vector-database
Experiment tracking: Weights & Biases and MLflow about 17 hours ago
help wanted deployment size: feature
dolev31/lasagna
20

Swap one RAG pipeline stage, hold the rest fixed, and measure whether the gain survives composition. The framework behind "Retrieval Augmented Generation Compositionality": 450 configurations, eight benchmarks, corpora from 5K to 200K documents.

Python
#ablation-study#agentic-rag#beir#benchmark#evaluation#hybrid-search#information-retrieval#langgraph#llm#mcp#mcp-server#python#rag#rag-evaluation#ragas#reproducibility#reranking#retrieval-augmented-generation#semantic-search#vector-database
Benchmark adapter: MIRACL (multilingual retrieval, 18 languages) about 17 hours ago
good first issue benchmark size: feature
dolev31/lasagna
20

Swap one RAG pipeline stage, hold the rest fixed, and measure whether the gain survives composition. The framework behind "Retrieval Augmented Generation Compositionality": 450 configurations, eight benchmarks, corpora from 5K to 200K documents.

Python
#ablation-study#agentic-rag#beir#benchmark#evaluation#hybrid-search#information-retrieval#langgraph#llm#mcp#mcp-server#python#rag#rag-evaluation#ragas#reproducibility#reranking#retrieval-augmented-generation#semantic-search#vector-database
Benchmark adapter: BRIGHT (reasoning-intensive retrieval) about 17 hours ago
good first issue benchmark size: feature
dolev31/lasagna
20

Swap one RAG pipeline stage, hold the rest fixed, and measure whether the gain survives composition. The framework behind "Retrieval Augmented Generation Compositionality": 450 configurations, eight benchmarks, corpora from 5K to 200K documents.

Python
#ablation-study#agentic-rag#beir#benchmark#evaluation#hybrid-search#information-retrieval#langgraph#llm#mcp#mcp-server#python#rag#rag-evaluation#ragas#reproducibility#reranking#retrieval-augmented-generation#semantic-search#vector-database
Research: how well does the judge agree with people? about 17 hours ago
help wanted research size: research
dolev31/lasagna
20

Swap one RAG pipeline stage, hold the rest fixed, and measure whether the gain survives composition. The framework behind "Retrieval Augmented Generation Compositionality": 450 configurations, eight benchmarks, corpora from 5K to 200K documents.

Python
#ablation-study#agentic-rag#beir#benchmark#evaluation#hybrid-search#information-retrieval#langgraph#llm#mcp#mcp-server#python#rag#rag-evaluation#ragas#reproducibility#reranking#retrieval-augmented-generation#semantic-search#vector-database
A benchmark in a Qdrant stack talks to Milvus, and fails about 17 hours ago
bug help wanted deployment size: feature
dolev31/lasagna
20

Swap one RAG pipeline stage, hold the rest fixed, and measure whether the gain survives composition. The framework behind "Retrieval Augmented Generation Compositionality": 450 configurations, eight benchmarks, corpora from 5K to 200K documents.

Python
#ablation-study#agentic-rag#beir#benchmark#evaluation#hybrid-search#information-retrieval#langgraph#llm#mcp#mcp-server#python#rag#rag-evaluation#ragas#reproducibility#reranking#retrieval-augmented-generation#semantic-search#vector-database
The knowledge graph lives in the memory of the process that built it, so served search never sees it about 17 hours ago
bug help wanted deployment stage:retriever knowledge-graph size: feature
dolev31/lasagna
20

Swap one RAG pipeline stage, hold the rest fixed, and measure whether the gain survives composition. The framework behind "Retrieval Augmented Generation Compositionality": 450 configurations, eight benchmarks, corpora from 5K to 200K documents.

Python
#ablation-study#agentic-rag#beir#benchmark#evaluation#hybrid-search#information-retrieval#langgraph#llm#mcp#mcp-server#python#rag#rag-evaluation#ragas#reproducibility#reranking#retrieval-augmented-generation#semantic-search#vector-database
Record what building the corpus costs: ingest-time LLM calls, tokens and time per arm about 17 hours ago
enhancement help wanted results knowledge-graph size: feature
dolev31/lasagna
20

Swap one RAG pipeline stage, hold the rest fixed, and measure whether the gain survives composition. The framework behind "Retrieval Augmented Generation Compositionality": 450 configurations, eight benchmarks, corpora from 5K to 200K documents.

Python
#ablation-study#agentic-rag#beir#benchmark#evaluation#hybrid-search#information-retrieval#langgraph#llm#mcp#mcp-server#python#rag#rag-evaluation#ragas#reproducibility#reranking#retrieval-augmented-generation#semantic-search#vector-database
Benchmark adapter: GraphRAG-Bench (when graphs help, by question type) about 17 hours ago
help wanted benchmark knowledge-graph size: feature
dolev31/lasagna
20

Swap one RAG pipeline stage, hold the rest fixed, and measure whether the gain survives composition. The framework behind "Retrieval Augmented Generation Compositionality": 450 configurations, eight benchmarks, corpora from 5K to 200K documents.

Python
#ablation-study#agentic-rag#beir#benchmark#evaluation#hybrid-search#information-retrieval#langgraph#llm#mcp#mcp-server#python#rag#rag-evaluation#ragas#reproducibility#reranking#retrieval-augmented-generation#semantic-search#vector-database
Record the compute each arm uses: CPU and GPU time, peak memory and index size about 17 hours ago
enhancement help wanted results knowledge-graph size: feature
dolev31/lasagna
20

Swap one RAG pipeline stage, hold the rest fixed, and measure whether the gain survives composition. The framework behind "Retrieval Augmented Generation Compositionality": 450 configurations, eight benchmarks, corpora from 5K to 200K documents.

Python
#ablation-study#agentic-rag#beir#benchmark#evaluation#hybrid-search#information-retrieval#langgraph#llm#mcp#mcp-server#python#rag#rag-evaluation#ragas#reproducibility#reranking#retrieval-augmented-generation#semantic-search#vector-database
Record what each built-in component's paper reports (claims) about 17 hours ago
good first issue size: feature
dolev31/lasagna
20

Swap one RAG pipeline stage, hold the rest fixed, and measure whether the gain survives composition. The framework behind "Retrieval Augmented Generation Compositionality": 450 configurations, eight benchmarks, corpora from 5K to 200K documents.

Python
#ablation-study#agentic-rag#beir#benchmark#evaluation#hybrid-search#information-retrieval#langgraph#llm#mcp#mcp-server#python#rag#rag-evaluation#ragas#reproducibility#reranking#retrieval-augmented-generation#semantic-search#vector-database
Vector backends: Chroma, LanceDB, Elasticsearch about 17 hours ago
help wanted deployment size: feature
dolev31/lasagna
20

Swap one RAG pipeline stage, hold the rest fixed, and measure whether the gain survives composition. The framework behind "Retrieval Augmented Generation Compositionality": 450 configurations, eight benchmarks, corpora from 5K to 200K documents.

Python
#ablation-study#agentic-rag#beir#benchmark#evaluation#hybrid-search#information-retrieval#langgraph#llm#mcp#mcp-server#python#rag#rag-evaluation#ragas#reproducibility#reranking#retrieval-augmented-generation#semantic-search#vector-database
An OpenAI-compatible answer endpoint, with streaming about 17 hours ago
help wanted deployment size: feature
dolev31/lasagna
20

Swap one RAG pipeline stage, hold the rest fixed, and measure whether the gain survives composition. The framework behind "Retrieval Augmented Generation Compositionality": 450 configurations, eight benchmarks, corpora from 5K to 200K documents.

Python
#ablation-study#agentic-rag#beir#benchmark#evaluation#hybrid-search#information-retrieval#langgraph#llm#mcp#mcp-server#python#rag#rag-evaluation#ragas#reproducibility#reranking#retrieval-augmented-generation#semantic-search#vector-database
Research: one hierarchical model of a whole sweep, from per-query scores about 17 hours ago
help wanted results research size: research
dolev31/lasagna
20

Swap one RAG pipeline stage, hold the rest fixed, and measure whether the gain survives composition. The framework behind "Retrieval Augmented Generation Compositionality": 450 configurations, eight benchmarks, corpora from 5K to 200K documents.

Python
#ablation-study#agentic-rag#beir#benchmark#evaluation#hybrid-search#information-retrieval#langgraph#llm#mcp#mcp-server#python#rag#rag-evaluation#ragas#reproducibility#reranking#retrieval-augmented-generation#semantic-search#vector-database
The knowledge graph on every surface: SDK, CLI, REST, MCP and the configurator about 17 hours ago
enhancement help wanted knowledge-graph size: feature
dolev31/lasagna
20

Swap one RAG pipeline stage, hold the rest fixed, and measure whether the gain survives composition. The framework behind "Retrieval Augmented Generation Compositionality": 450 configurations, eight benchmarks, corpora from 5K to 200K documents.

Python
#ablation-study#agentic-rag#beir#benchmark#evaluation#hybrid-search#information-retrieval#langgraph#llm#mcp#mcp-server#python#rag#rag-evaluation#ragas#reproducibility#reranking#retrieval-augmented-generation#semantic-search#vector-database
Knowledge graph explorer in the configurator: see, iterate on and measure the graph about 17 hours ago
accessibility enhancement help wanted knowledge-graph size: feature
dolev31/lasagna
20

Swap one RAG pipeline stage, hold the rest fixed, and measure whether the gain survives composition. The framework behind "Retrieval Augmented Generation Compositionality": 450 configurations, eight benchmarks, corpora from 5K to 200K documents.

Python
#ablation-study#agentic-rag#beir#benchmark#evaluation#hybrid-search#information-retrieval#langgraph#llm#mcp#mcp-server#python#rag#rag-evaluation#ragas#reproducibility#reranking#retrieval-augmented-generation#semantic-search#vector-database
Study: what the knowledge graph contributes, on data, with quality, time, compute and cost about 17 hours ago
help wanted results research knowledge-graph size: research
dolev31/lasagna
20

Swap one RAG pipeline stage, hold the rest fixed, and measure whether the gain survives composition. The framework behind "Retrieval Augmented Generation Compositionality": 450 configurations, eight benchmarks, corpora from 5K to 200K documents.

Python
#ablation-study#agentic-rag#beir#benchmark#evaluation#hybrid-search#information-retrieval#langgraph#llm#mcp#mcp-server#python#rag#rag-evaluation#ragas#reproducibility#reranking#retrieval-augmented-generation#semantic-search#vector-database
The image's Qdrant client is seven minor versions ahead of the stack's Qdrant server about 18 hours ago
bug good first issue deployment
dolev31/lasagna
20

Swap one RAG pipeline stage, hold the rest fixed, and measure whether the gain survives composition. The framework behind "Retrieval Augmented Generation Compositionality": 450 configurations, eight benchmarks, corpora from 5K to 200K documents.

Python
#ablation-study#agentic-rag#beir#benchmark#evaluation#hybrid-search#information-retrieval#langgraph#llm#mcp#mcp-server#python#rag#rag-evaluation#ragas#reproducibility#reranking#retrieval-augmented-generation#semantic-search#vector-database
Knowledge-graph RAG: the plan, from a flag to methods you can compare 1 day ago
help wanted knowledge-graph
dolev31/lasagna
20

Swap one RAG pipeline stage, hold the rest fixed, and measure whether the gain survives composition. The framework behind "Retrieval Augmented Generation Compositionality": 450 configurations, eight benchmarks, corpora from 5K to 200K documents.

Python
#ablation-study#agentic-rag#beir#benchmark#evaluation#hybrid-search#information-retrieval#langgraph#llm#mcp#mcp-server#python#rag#rag-evaluation#ragas#reproducibility#reranking#retrieval-augmented-generation#semantic-search#vector-database
The stress suite looks for the Compose files in a directory that does not exist 3 days ago
bug good first issue
dolev31/lasagna
20

Swap one RAG pipeline stage, hold the rest fixed, and measure whether the gain survives composition. The framework behind "Retrieval Augmented Generation Compositionality": 450 configurations, eight benchmarks, corpora from 5K to 200K documents.

Python
#ablation-study#agentic-rag#beir#benchmark#evaluation#hybrid-search#information-retrieval#langgraph#llm#mcp#mcp-server#python#rag#rag-evaluation#ragas#reproducibility#reranking#retrieval-augmented-generation#semantic-search#vector-database
The CLI reference page changes with the installed typer version 3 days ago
bug good first issue
dolev31/lasagna
20

Swap one RAG pipeline stage, hold the rest fixed, and measure whether the gain survives composition. The framework behind "Retrieval Augmented Generation Compositionality": 450 configurations, eight benchmarks, corpora from 5K to 200K documents.

Python
#ablation-study#agentic-rag#beir#benchmark#evaluation#hybrid-search#information-retrieval#langgraph#llm#mcp#mcp-server#python#rag#rag-evaluation#ragas#reproducibility#reranking#retrieval-augmented-generation#semantic-search#vector-database
Two services starting together race to create the collection, and one fails to start 7 days ago
bug good first issue deployment
dolev31/lasagna
20

Swap one RAG pipeline stage, hold the rest fixed, and measure whether the gain survives composition. The framework behind "Retrieval Augmented Generation Compositionality": 450 configurations, eight benchmarks, corpora from 5K to 200K documents.

Python
#ablation-study#agentic-rag#beir#benchmark#evaluation#hybrid-search#information-retrieval#langgraph#llm#mcp#mcp-server#python#rag#rag-evaluation#ragas#reproducibility#reranking#retrieval-augmented-generation#semantic-search#vector-database
Benchmark adapter: 2WikiMultihopQA, with its distractor paragraphs 7 days ago
help wanted benchmark
dolev31/lasagna
20

Swap one RAG pipeline stage, hold the rest fixed, and measure whether the gain survives composition. The framework behind "Retrieval Augmented Generation Compositionality": 450 configurations, eight benchmarks, corpora from 5K to 200K documents.

Python
#ablation-study#agentic-rag#beir#benchmark#evaluation#hybrid-search#information-retrieval#langgraph#llm#mcp#mcp-server#python#rag#rag-evaluation#ragas#reproducibility#reranking#retrieval-augmented-generation#semantic-search#vector-database
`lasagna component validate` rejects an `agentic_pipeline` 9 days ago
good first issue stage:agentic_pipeline
dolev31/lasagna
20

Swap one RAG pipeline stage, hold the rest fixed, and measure whether the gain survives composition. The framework behind "Retrieval Augmented Generation Compositionality": 450 configurations, eight benchmarks, corpora from 5K to 200K documents.

Python
#ablation-study#agentic-rag#beir#benchmark#evaluation#hybrid-search#information-retrieval#langgraph#llm#mcp#mcp-server#python#rag#rag-evaluation#ragas#reproducibility#reranking#retrieval-augmented-generation#semantic-search#vector-database