- Title / ID
- Publication year
- Document type
- Language
- License
- Country/region
- PDF URL
Forget GraphRAG — A 4B AI does the work NOW
GraphRAG 系统普遍建错图 — 5 大问题导致"科学结构在 retrieval 阶段被破坏"，AI agent 因此失败。K1 (Agent Native Knowledge Orchestration) 解决方案：(1) 5 schema 多模态 KG · (2) 4B 专用 extraction LLM · (3) 3 源 retrieval（不用 vector DB） · (4) 17 MCP tools + CLI。
关键 insight：4B 模型 + SOTA reasoning model 分离 = "less money less expensive"。
| # | 问题 | 影响 | 例子 |
|---|---|---|---|
| 1 | Flatten scientific structure | 条件结构消失，依赖关系丢失 | "improves the robustness only when X and Y" → 变成 "improves robustness" |
| 2 | Ignores multimodal | table / figure 数据全丢 | table 3 / figure 3 的非文本数据无法 retrieve |
| 3 | Lacks abstractions | 只取 metadata，缺科学抽象 | 有 dataset/method/metric，没有 motivation/contribution/limitation |
| 4 | Citations as flat edges | 丢失引用语义 | "A cites B" → 没区分 support/contradict/extend/benchmark/criticize/reuse |
| 5 | Retrieves only chunks | 还是 text retrieval | 拿不到 benchmark results / failure claims / equations / lineage |
"Our latest research AI agents fail not because they are stupid or have not enough complex sub-agent and multi-agent structure but because the scientific structure / information / knowledge itself gets destroyed during the retrieval."
解析 PDF → 5 schemas 结构化
专用 expert model（不是 chat / reasoning）
17 MCP tools + Python package + Unified CLI
| 字段 | 提取值 |
|---|---|
| Method | Socratic Policy Optimization |
| What it improves | RL convergence |
| Mechanism | Teacher-guided curriculum generation |
| Baseline | PPO (2017) |
| Problem | Exploration inefficiency |
| Operator | 用途 | 例子 |
|---|---|---|
| Seed resolution | 解析 query → 实体 | "Socratic PPO" → 论文名 + 作者 + 方法 + 引用 |
| Citation lineage | 演进链 | "What led to this methodology?" → 遍历 graph |
| Limitation retrieval | 未解决问题 | "Which limitations remain?" |
| Comparative retrieval | 文献对比 | "What beats classical PPO?" → benchmark + delta + failure |
"Graph traversal... is NOT vector building or vector databases or vector searching — this is reasoning over the graph paths and elements here that we need for the long reasoning task."
"Think about the nature of the job. What we have to do is extraction and even more structured extraction optimization and not simple reasoning. So what we need is here a massive alignment to our new schema."
| Reward | 作用 |
|---|---|
| Format reward | 输出符合 schema 格式 |
| JSON reward | JSON 合法 + 结构化 |
| Task reward | NER F1 + Relation Extraction + Long-form structured |
"Socratic PO is a method, PPO is the baseline, this is a dataset"
extend / compare / reuse / adapt / contrast / generalize / modify / criticize
method / role / evidence / context / analysis
4B 模型做 extraction · SOTA 模型做 reasoning · 避免 giant expensive LLM 用于所有 task
"Less money less expensive."
在所有 3 个 multihop QA benchmark 上，K1 都 outperforms 所有 baseline（包括 LightRAG / HippoRAG / GFMRAG / HippoRAG 2）
注：Naked GPT-5.2 在 knowledge / research question 表现类似 41.8%
核心关联 — K1 是 hybrid-llm-router 的 production-grade 实现：
4B 模型本地能跑 → 本地 + cloud reasoning 的最优分工
Graph traversal reasoning 跟 IndyDevDan "8K context limit" 完全互补
OpenSwarm = 多 agent 协作 · K1 = single agent with 3-source retrieval
两条路线可结合：K1 节点 + OpenSwarm 协调
Sam "inference underinvested" → K1 = inference 优化实例
"Always fresh data" → Sam 3 forks 之一（compute + data shortage）
"less money less expensive" = Sam "80% democratization" 具体路径
DataCamp 学 RAG → K1 是 RAG 2.0 范式（无 vector DB）
Multimodal + structured extraction 是进阶技能
K1 论文 6/11 刚出 → 7 个 MCP tools 立即可用 → 跟 hybrid-llm-router skill 完美互补
这是 RAG 2.0 范式：从 vector DB → 3 源 retrieval（web + KG + traversal）