Method overview

CausalRAG answers a question over an entity-level text graph (causalrag/method.py::query_causal_rag), following the paper:

1. Seed retrieval (Sec. 4.2). The query is embedded with all-MiniLM-L6-v2 and matched against the graph nodes via a FAISS IndexFlatL2; the top-k nodes become the seeds.

2. s-hop expansion (Sec. 4.2). Starting from the seeds, the retriever walks s hops along the graph's relationships, collecting the traversed cause -> effect edges into a causal subgraph.

3. Causal-discovery report (Sec. 4.2). The subgraph is rendered as "Network Data" and an LLM turns it into a structured causality analysis report using the Causal Discovery Prompt (Appendix A.1 / Figure 7).

4. Grounded answer (Sec. 4.3). A second LLM call merges that report with the query into the final answer using the Causal Summary Prompt (Appendix A.2 / Figure 7).

The return value is a dict with answer, causal_report, network_data, seed_nodes, and a meta block (num_relationships, k, s, model).

Key parameters

Parameter	Meaning	Default
model	discovery / summary LLM	gpt-4o-mini
embed_model	sentence-transformers model for retrieval	all-MiniLM-L6-v2
k	number of initial seed nodes	3
s	number of expansion hops	3
response_type	target length/format for the answer	Multiple Paragraphs

Both model and embed_model match the paper's Appendix C. k = s = 3 is the main experimental setting; the parameter study sweeps k, s ∈ {1, ..., 5}.