Take citations away from the model: dev open-sources sourced-quote MCP server for 50 founders
Otherwise_Vehicle75 · reddit · 2026-10-26
The author found that asking LLMs to cite sources makes fabricated quotes look more convincing. The fix: the model never writes a source itself — it calls a search tool where every passage comes back with title, date and URL attached by the index, so it can't invent provenance.
Built for 50 founders/investors (20,800 passages), open source (MIT), works as an MCP server in Claude, ChatGPT, Cursor and Codex. Lessons:
- Source curation beat retrieval tuning: an agent shortlisting the 15-45 genuinely relevant essays/talks per person mattered more than any search tweak.
- Plain BM25 was enough: a few hundred passages per person on one domain; 80 lines of code, no vector DB. Weak on vocabulary mismatch — the model just re-searches with different words.
- Show disagreement instead of averaging it: sourced passages reveal that Pieter Levels says don't quit until side income beats your job, while others say the opposite — the disagreement is the useful part.
- Index short third-person notes, not raw transcripts: a preprocessing pass into sectioned notes with verbatim quotes made results far cleaner.
Still unsolved: cheap citation-faithfulness evals without hand-labeling.
More from coding & agent
- Just feed it to Codex: 10 good variations in 5 minutes — var_epsilon · 2026-10-26
- Local book-to-audiobook-video pipeline runs on 8GB VRAM with Qwen Image + TTS — top115 · 2026-10-26
- CLI coding harnesses silo session data per tool — a dev wants cross-harness sync — scubanarc · 2026-10-26
- Zero to One is cliché—until frontier labs and Claude Code homogenize everyone's thinking — akbirthko · 2026-10-26
- A decoy MCP server got hit by 7 AI tools in 7 minutes — 4 asked for the full tool list — KnowledgeOk7634 · 2026-10-26
- Do developers actually use multi-agent orchestration, or is one agent enough? — latent_lurker8587 · 2026-10-26