AgentSkillsCN

rag-proj

用于 NIST-PQC-RAG-Agent 仓库的规划、实施、调试与评估：包括从 NIST PQC PDF 文件中完成数据摄取与分块处理，采用 BM25 与向量检索相结合并进行融合的混合检索方案，借助 LangChain/LangGraph 实现智能代理控制，并构建具备可量化增益的评估框架。

SKILL.md
--- frontmatter
name: rag-proj
description: >
  Use for planning, implementing, debugging, and evaluating the nist-pqc-rag-agent repo:
  ingestion/chunking over NIST PQC PDFs, hybrid retrieval (BM25+vector) with fusion,
  LangChain/LangGraph agent control, and an evaluation harness with measurable deltas.

Purpose

This skill turns “build a RAG agent” into a week-long, recruiter-legible mini production system: deterministic pipeline, citation-first answers, bounded agent loops, and eval-driven iteration.

Always do first (no exceptions)

  1. •Read project_overview.md to anchor scope, milestones, and contracts.
  2. •Inspect current repo structure (don’t guess filenames).
  3. •Identify the smallest change that advances the current milestone.

Non-negotiables (PQC PDFs are prickly)

  • •Preserve page numbers and page spans for citations.
  • •Chunk sizes should stay focused (~250–400 tokens, overlap OK). Avoid giant “muddy vectors”.
  • •Treat tables/algorithms/math blocks as first-class citizens: don’t destroy structure if avoidable.

System design rules (employer-appealing)

  • •Retrieval must be swappable: app code depends on Retriever interface, not FAISS details.
  • •Default retrieval: hybrid BM25 + vector + Reciprocal Rank Fusion (RRF).
  • •Agent must be bounded: step budget + tool-call budget + stop rules.

What to output for any request

When the user asks you to do something:

  • •A short “success criteria” statement for this change
  • •The exact files to create/modify
  • •Copy/paste-ready code blocks per file
  • •The exact commands to run to verify
  • •If the change claims improvement: include eval deltas (before/after)

Common workflows

A) Planning (Day-level)

  • •Map work to Week-1 milestones in project_overview.md
  • •Produce a checklist with acceptance criteria and the one metric that should move

B) Implementing

  • •Make minimal diffs
  • •Add/adjust the smallest test or smoke check
  • •Keep JSONL schema stable

C) Debugging

  • •Reproduce from logs
  • •Localize root cause
  • •Patch + add regression check

D) Evaluation

  • •Run harness
  • •Report retrieval metrics + citation compliance
  • •Only then propose next improvement

Handy commands (adapt to what exists in repo)

  • •Tree: python .agents/skills/rag_proj/scripts/smoke.py --tree
  • •Smoke checks: python .agents/skills/rag_proj/scripts/smoke.py
  • •Optional run mode: python .agents/skills/rag_proj/scripts/smoke.py --run