What it solves
One broad query can under-retrieve when it contains multiple constraints, time periods, or entities. Decomposition separates the work into smaller retrieval targets.
01 · ReadUnderstand the original question.
02 · SplitGenerate focused sub-queries.
03 · RetrieveRun each path concurrently.
04 · RejoinDeduplicate and cite context.
How Mneme handles it
The configured LLM endpoint proposes sub-queries. Local retrieval searches the resulting terms and meanings, bounds the context, and keeps source metadata attached.
What you can inspect
- The original question and its source markers.
- Bounded retrieved context rather than a document dump.
- Exact files and chunks behind the final answer.
Result before terminology
Decomposition helps cover compound questions; it does not promise that every answer is correct. Inspect the cited trail.
Try it
python -m src.graph_rag \
--files /path/to/docs \
--query "What are the main findings?"