LLM-RE-KnowledgeBase-Agent - GitLite

KO | EN

gitlite — search

#swift #llm #anthropic #cross-platform #windows #vue #shell #mcp #nodejs #terminal

LLM-RE-KnowledgeBase-Agent

(★ 9)

LLM-driven reverse engineering knowledge base custodian — ingests blogs, pentest findings, and malware write-ups into a structured, anti-hallucination markdown wiki with an interactive knowledge graph.

파일 탐색기

ZIP 다운로드

- - _common-metadata.md
  - analysis-notes.md
  - api-reference.md
  - blog-writeup.md
  - malware-analysis.md
  - pentest-finding.md
  - research-paper.md
- - crawl.py
  - graph_generator.py
  - requirements.txt
- core-rules.md
- index-formats.md
- ingest-protocol.md
- maintenance.md
- read-protocol.md
- write-pipeline.md
- architecture.png
- architecture.svg
- - index.html
AGENTS.md
README.md

# CDN으로 사용하기

jsDelivr

jsDelivr는 공개 GitHub 리포지토리를 별도 설정 없이 CDN으로 즉시 서빙합니다. 버전과 파일을 고르면 웹페이지에 바로 붙일 수 있는 링크와 예시 코드가 만들어집니다.

버전

파일

링크

예시

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

> A file-based, LLM-driven **Reverse Engineering Knowledge Base custodian**. It ingests blogs, papers, pentest findings, and malware write-ups, compiles them into a rigorously structured, cross-linked markdown wiki with anti-hallucination provenance markers, and renders the whole thing as an interactive knowledge graph.

The agent is not a chatbot wrapper — it is a set of instruction files (`AGENTS.md` + `.kb/`) that turn any capable coding/agent LLM (Claude Code, Cursor, etc.) into a disciplined librarian for security-research knowledge. All state lives in plain markdown on disk, so the KB is portable, diff-able, and readable by other agents via direct file traversal.

---

## Why this exists

RE and security notes rot fast: opcodes get paraphrased away, tool versions go unpinned, blog claims get promoted to "fact," and a year later you can't tell what you *verified* from what an LLM *guessed*. This agent enforces the opposite discipline:

- **Zero-loss compilation** — offsets, hex, registers, JNI signatures, and bypass scripts are preserved verbatim, never summarized.
- **Provenance on every claim** — inline markers separate verified facts (`[V]`) from inference (`[I]`), unverified secondary sources (`[U]`), version-specific behavior (`[VS]`), contradictions (`[X]`), and pure model reasoning (`[AI Synthesis]`).
- **Navigable by machines** — every article carries YAML frontmatter and lives in a hub-and-chunk graph so *other* LLMs can traverse it without confabulating.

---

## Architecture

<p align="center">
  <img src="docs/architecture.png" alt="3-tier AI agent system: user input commands (ingest / compile / search) → AI agent orchestrator (rules + dispatcher) → markdown knowledge vault (raw feeds → compiled graph nodes)" width="620">
</p>

The system is three tiers: **user commands** (`ingest` / `compile` / `search`) drive an **AI agent orchestrator** (behavioral rules + a mode dispatcher), which reads from and writes to the **markdown knowledge vault** — turning raw inbox feeds into a compiled, cross-linked graph of articles.

```
./ (project root)
├── AGENTS.md                  ← Master orchestrator / mode dispatcher (read every session)
├── .kb/                       ← The agent's "firmware": rules, protocols, schemas, scripts
│   ├── core-rules.md          ← Anti-hallucination system + behavioral rules (always loaded)
│   ├── read-protocol.md       ← Query mode: how to traverse the KB and answer
│   ├── write-pipeline.md      ← Compile mode: raw file → structured wiki articles
│   ├── ingest-protocol.md     ← Crawl mode: URL → raw markdown + local images
│   ├── index-formats.md       ← Templates for master/topic/cross-topic indexes
│   ├── maintenance.md         ← health-check / gap report / contradictions
│   ├── schemas/               ← Per-content-type article schemas
│   │   ├── _common-metadata.md
│   │   ├── pentest-finding.md
│   │   ├── malware-analysis.md
│   │   ├── research-paper.md
│   │   ├── api-reference.md
│   │   ├── analysis-notes.md
│   │   └── blog-writeup.md
│   └── scripts/
│       ├── crawl.py           ← Portable web crawler (readability + image downloader)
│       ├── graph_generator.py ← Builds graph_data.js from wiki links
│       └── requirements.txt
└── knowledge_base/            ← The isolated KB environment (all persistent state)
    ├── raw/                   ← Inbox: unprocessed feeds/reports + downloaded assets
    ├── wiki/                  ← The compiled markdown KB (Obsidian-style [[links]])
    └── graph/                 ← Interactive knowledge-graph visualizer (index.html)
```

The LLM reads `AGENTS.md` first, matches the user's request to a **mode**, and loads only the instruction files that mode needs.

---

## Core workflow

```
ingest <url>   →   compile <file>   →   search <query>
   crawl            structure             answer
   ─────            ─────────             ──────
 raw/feeds/*     wiki/**/*.md +        unified technical
 raw/assets/*    updated indexes +     response with inline
                 regenerated graph     provenance markers
```

### 1. Ingest — crawl a URL into the inbox
`crawl.py` fetches a page with a realistic User-Agent, extracts the main article via `readability`, recovers `<figure>`/SVG diagrams that readability strips, downloads every image locally (converting `.webp` → `.png`), and writes clean markdown to `knowledge_base/raw/feeds/<slug>.md`.

### 2. Compile — turn raw notes into a structured wiki
The agent classifies the content type (pentest finding, malware analysis, research paper, API doc, analysis notes, or blog), loads the matching schema, and splits deep dives into a **hub + semantic chunks** (one chunk per *technical mechanism*, never per product name). It writes YAML frontmatter, cross-links every article, updates topic and cross-topic indexes, verifies all `[[wiki links]]` resolve, and logs the operation.

### 3. Search — query the KB
The agent routes through `_master-index.md` → topic `_index.md` → article, then composes **one seamless, highly granular technical document** — exact function names, opcodes, structs, and code — with every sentence tagged by provenance. If something isn't in the KB, it says so explicitly rather than blending in training-data priors.

### 4. Visualize — interactive knowledge graph
`graph_generator.py` parses every wiki file, extracts `[[wiki/...]]` links, builds a node/edge graph colored by topic, and writes `knowledge_base/graph/graph_data.js`. Open `knowledge_base/graph/index.html` for a force-directed, searchable, clickable map of the KB.

---

## Command reference

Drive the agent with natural language or these shorthand commands (see `AGENTS.md`):

| Command | Action |
|---|---|
| `ingest <url>` | Crawl a URL → raw markdown + local images inbox |
| `compile <file>` | Process an unprocessed `raw/` file into structured wiki articles |
| `search <query>` | Query the KB → unified answer (KB facts + AI synthesis, inline) |
| `graph` | Build & open the interactive knowledge-graph visualizer |
| `health-check` | Scan for dead links, orphans, missing YAML, stale stubs |
| `gap report` | Report dead links, stale stubs, and underdeveloped topics |
| `contradictions` | List all articles carrying `[X]` conflict markers |

---

## Provenance marker system

Every factual claim in a wiki article carries a marker so downstream LLMs never confuse sources:

| Marker | Meaning |
|---|---|
| `[V]` | **Verified** — directly confirmed from the cited primary source |
| `[I]` | **Inferred** — logical conclusion from verified facts |
| `[U]` | **Unverified** — from a blog/forum/secondary source |
| `[VS: Frida 16.x]` | **Version-specific** — only applies to the stated version/environment |
| `[X: see <article>]` | **Contradicted** — conflicts with another article |
| `[AI Synthesis]` | **Model intelligence** — general pre-trained reasoning or code patterns |

Every article must also declare `## What This Article Does NOT Cover` and `## Open Questions / Unverified Claims` to keep scope boundaries explicit.

---

## Getting started

### Prerequisites
- Python 3.10+ (the scripts use `X | None` type-union syntax)
- An agent/LLM host that reads project instruction files (e.g. Claude Code)

### 1. Clone
```bash
git clone https://github.com/fatalSec/LLM-RE-KnowledgeBase-Agent.git
cd LLM-RE-KnowledgeBase-Agent
```

### 2. Set up the Python environment (for crawling & graphing)
```bash
python3 -m venv .kb/scripts/.venv
.kb/scripts/.venv/bin/pip install -r .kb/scripts/requirements.txt
```

### 3. Crawl a page into the inbox
```bash
.kb/scripts/.venv/bin/python .kb/scripts/crawl.py "https://example.com/some-re-writeup"
```

### 4. Point your agent at the repo
Open the project in your agent host and ask it to, for example, `compile <slug>.md`, `search "DexProtector anti-frida"`, or run a `health-check`. The agent reads `AGENTS.md` and takes it from there.

### 5. Explore the graph
```bash
.kb/scripts/.venv/bin/python .kb/scripts/graph_generator.py --open
```

---

## Scripts

| Script | Purpose | Key dependencies |
|---|---|---|
| `.kb/scripts/crawl.py` | Fetch a URL, extract the main article, download images locally, emit raw markdown. Handles data-URI images, SVG/figure recovery, and `.webp`→`.png` conversion. | `requests`, `beautifulsoup4`, `lxml`, `readability-lxml`, `markdownify`, `Pillow` |
| `.kb/scripts/graph_generator.py` | Parse the wiki, extract Obsidian links, build the node/edge graph, write `graph_data.js`, optionally open the visualizer. | `PyYAML` (optional; falls back to regex frontmatter parsing) |

> **Note:** `crawl.py` prefers the native macOS `sips` tool for `.webp` conversion and only falls back to `Pillow` when `sips` is unavailable. On non-macOS systems, `Pillow` handles the conversion.

The knowledge-graph visualizer (`knowledge_base/graph/index.html`) loads `force-graph` and `marked` from CDNs and expects internet access when opened in a browser.

---

## Design principles

- **Files, not a database.** The entire KB is markdown + YAML on disk — portable, versionable, and readable by any agent or human.
- **Mechanisms are the unit of knowledge.** Articles are split by technical mechanism, not by product or target name.
- **No orphans.** Every article is indexed in its topic and linked from at least one other article.
- **Version-pin everything.** Tool and SDK versions are always explicit (`Frida 16.x` vs `15.x`, `Android 13` vs `14`).
- **Verify, then trust.** A mandatory verification loop re-checks outputs against source so no opcode or detail is silently dropped.

---

## License

No license file is currently included. Add one before publishing if you intend to permit reuse.