Agentic CustomNerd

Agentic RAG backend for CloudNerd (Stack Overflow) and DietNerd (PubMed).
Same pipeline for both — switch with AGENT_DOMAIN.

This repo is the agent layer only. It reuses retrieval / relevance / synthesis from a sibling CustomNerd linear backend (customnerd-backend/).

Layout (expected)

your-workspace/
├── agentic-customnerd/          # this repo (or rename to cloudnerd-agent-backend)
└── customnerd-backend/          # linear CustomNerd backend (clone separately)
    └── saved_states/
        ├── CloudNerd/
        └── DietNerd/

config.py resolves the legacy root as ../customnerd-backend next to this folder.

Pair with CustomNerd

1. Clone / place the linear backend as a sibling folder named customnerd-backend.
2. Ensure domain packs exist under customnerd-backend/saved_states/CloudNerd and DietNerd (prompts + search APIs).
3. Put API keys in customnerd-backend/variables.env (or copy variables.env.template → variables.env here for agent overrides).
4. Prefer reusing the linear backend venv (it already has BioPython, OpenAI, etc.).

Domains (AGENT_DOMAIN)

Domain	Retrieval	Notes
CloudNerd	SO cascade (RETRIEVAL_MODE=cascade)	Needs STACK_API_KEY
DietNerd	PubMed Entrez (RETRIEVAL_MODE=legacy)	Needs ENTREZ_EMAIL (+ NCBI_API_KEY)

On boot (and via POST /agent/load_domain), the agent copies
customnerd-backend/saved_states/<DOMAIN>/ into the live backend and reloads prompts / search.

# DietNerd
AGENT_DOMAIN=DietNerd
RETRIEVAL_MODE=legacy

# CloudNerd
AGENT_DOMAIN=CloudNerd
RETRIEVAL_MODE=cascade

Hot-switch: POST /agent/load_domain with form field domain=DietNerd.

Modes (AGENT_MODE)

Mode	Behavior
hybrid (default)	Enhanced pipeline + optional rescue pass
parity	Enhanced pipeline, no rescue
adaptive	ReAct or LangGraph (USE_LANGGRAPH=true)

Pipeline (high level)

search → organize → classify → assess
  → (optional re-search)
  → pre-verify faith → synthesize → claim filter → post-verify
  → (optional hybrid rescue) → answer + citations

Quick start

# From parent workspace that also contains customnerd-backend/
cd agentic-customnerd
copy variables.env.template variables.env
# edit AGENT_DOMAIN / keys as needed

..\customnerd-backend\venv\Scripts\python.exe -m uvicorn main:app --host 127.0.0.1 --port 8001

Health: GET http://127.0.0.1:8001/health

Eval (from linear backend)

cd ..\customnerd-backend
.\venv\Scripts\python.exe eval_rag.py --csv path\to\input.csv `
  --api-url http://127.0.0.1:8001 --output ..\result\agent_eval --agent-trace

Key env vars

Variable	Default	Description
AGENT_DOMAIN	CloudNerd / DietNerd	Domain pack to activate
AGENT_MODE	hybrid	hybrid, parity, or adaptive
AGENT_FAST_ORGANIZE	true	Fast article map (0 LLM per paper for Diet/SO)
AGENT_PREFER_COMBINED_SYNTHESIS	true	Grounded combined_lite answers
AGENT_FAITH_FIRST	true	Faithfulness-oriented verify / regen
HYBRID_RESCUE_ENABLED	true	Rescue pass in hybrid mode
RETRIEVAL_MODE	legacy / cascade	Diet vs Cloud retrieval

Do not commit variables.env (secrets). Use variables.env.template only.