Multi-Agentic Research Platform

MARP exists because a single LLM call cannot be audited: it answers, but it cannot show its work. The platform decomposes research into five specialized agents in sequence, iterates until the answer meets a confidence threshold or hits the iteration cap, and returns every step traced and timed for inspection.

The Planner turns the question into a structured retrieval plan (typed PlanStep objects: sub-question plus search query). The Retriever runs cosine-similarity search against PostgreSQL with pgvector — embeddings generated through Gemini's embedContent API — and returns ranked chunks with source metadata and similarity scores. The Writer drafts from evidence, the Critic challenges the draft, and the Verifier checks claims before release; the Critic→Writer loop repeats until confidence clears the bar. Every agent emits typed trace events.

• evidence grounding — no claim ships without a retrieval trail behind it
• bounded iteration — the critique loop must converge or stop at a hard cap, never spin
• LLM output fragility — structured JSON from a model cannot be assumed valid

• five single-responsibility agents over one omnibus prompt: more inference calls per question, but each stage emits typed traces and can be replaced without retraining the others
• loop-until-confident over single-pass answers: response latency deliberately spent on claim-level verification, bounded by a hard iteration cap so the spend cannot run away
• pgvector inside Postgres over a managed vector service: one database, one operational surface, one failure domain to observe

• the Planner's JSON parsing can fail on malformed LLM output — it degrades to treating the raw output as a single search query rather than aborting the run
• the Retriever returns an empty list gracefully when the vector store has nothing — downstream stages handle absence of evidence as a first-class state
• the Retriever currently executes only the first PlanStep of a multi-step plan — a known limit, preserved in the trace rather than papered over

python · postgres + pgvector · gemini embeddings · typescript · docker

The interesting problem was never the model — it was how intelligence behaves under constraints: what pipeline shape makes an LLM's answer auditable instead of plausible. Single-responsibility stages with typed contracts and traces are boring, and boring is what can be debugged.

• > execute the full retrieval plan, not just its first step
• > confidence calibration against held-out questions