Looking for a local "NotebookLM for lawyers" setup – what am I doing wrong?

In three linesA lawyer seeks to build a local RAG system to analyze case files (correspondence, contracts, court decisions) with citations. After testing Qwen 3.5 9B and gpt-oss-20b via LM Studio + Big RAG, he encounters two issues: insufficient speed (~2.2 tok/s) and model refusal to cite his own documents, generating generic explanations instead of context-grounded analysis.

Summary generated by Claude — human-verified