---

💡 Why Super RAG?

Enterprise security copilots (Microsoft Security Copilot, CrowdStrike Charlotte) live in the cloud, cost a fortune, and are built for defenders. Super RAG flips that:

• 🔒 Fully offline — runs against a local LLM in LM Studio. Perfect for air-gapped labs, sensitive engagements, and regions with restricted cloud access.
• ⚔️ Built for offensive reasoning — not just "summarize this alert," but "I see Apache 2.4.49 — what's my next move?" and it chains recon → vuln-ID → exploitation.
• 📚 Grounded in real knowledge — indexes 1,110 hand-curated notes plus HackTricks, PayloadsAllTheThings, the OWASP cheat sheets, and hundreds of CTF write-ups. Every answer is cited back to its source file.
• 🧩 Plugs into your agents — exposed as an MCP server, so Hermes, OpenClaw, Claude Desktop, or any MCP client gains a cybersec_search tool instantly.

⚠️ For authorized use only. This is a tool for pentesters, CTF players, and security researchers operating with explicit written permission. See Responsible Use.

---

✨ Features

🔎 13-strategy hybrid retrieval	Dense vectors + full-text BM25, fused with RRF, then multi-hop, corrective, context-aware and re-ranking passes — tuned per pentest phase.
🤖 Agentic pentest loop	reason → act → observe → reflect, with stuck-loop detection and a hybrid mode (auto-recon, manual approval before exploitation).
🍯 Evasion awareness	Built-in honeypot, WAF (8 vendors), and firewall detection — so the agent doesn't waste moves on a tarpit.
🛡️ Hallucination guard	A command registry validates every tool invocation against known-good flags before anything runs.
📝 Report generation	One command turns findings into a professional pentest report or a HackerOne-style bug-bounty submission, with CVSS and evidence.
🔌 MCP integration	One shared server, many agents — cybersec_search, cybersec_answer, cybersec_status.
🎓 Fine-tune ready	Extracts real HTB/VulnHub write-up reasoning into JSONL (never fabricated scenarios).

---

🏗️ Architecture

flowchart LR
    subgraph Knowledge["📚 Knowledge (61,552 chunks)"]
        V["1,110 curated notes"]
        D["HackTricks · PayloadsAllTheThings · OWASP"]
        W["CTF / HTB write-ups"]
    end
    subgraph Engine["🧠 Super RAG"]
        I["ingest.py<br/>chunk + embed (parallel)"]
        Q[("Qdrant<br/>hybrid index")]
        R["rag_engine.py<br/>13 RAG strategies"]
    end
    subgraph Local["💻 LM Studio (localhost:1234)"]
        E["nomic-embed-text"]
        L["gpt-oss-20b"]
    end
    A["agent.py<br/>pentest loop"]
    M["rag_mcp.py<br/>MCP server :8765"]

    V & D & W --> I --> Q
    I -.embeddings.-> E
    Q --> R --> L
    R --> A
    R --> M
    M --> Hermes & OpenClaw & Claude["Claude Desktop"]

Three tiers, depth over scale: embedded Qdrant (no Docker) for hybrid search → 13 combined RAG strategies for context assembly → a local LLM for generation.

---

🔬 The 13 RAG strategies (combined on every query)

Most projects use #1 and wonder why retrieval is mediocre. Super RAG layers 13, each earning its place in a pentest workflow:

#	Strategy	What it buys you
4	Hybrid (vector + BM25, RRF-fused)	Semantic recall and exact-string recall for CVE-2024-1086, --no-preauth, SeDebugPrivilege
17	Multi-Hop	Port 389 → LDAP enum → user list → AS-REP roast → hash → crack, each hop informed by the last
9	Agentic	The agent decides when and what to retrieve mid-engagement
6	Memory-Augmented	Remembers what was tried hours ago, so it never re-runs a dead path
3	Corrective	Detects weak retrieval and re-queries with reformulated terms
8	Context-Aware	Filters to the target environment (Windows/AD vs web vs cloud)
18	Reasoning re-rank	Keyword-overlap boost so the most useful chunk floats up, not just the most similar
13	Adaptive	Broad scope during recon, narrow and precise during exploitation
21	Hierarchical	General → specific drill-down, mirroring the ATT&CK structure
5	Speculative	Pre-fetches likely follow-ups in the background to cut latency
11	Self-RAG	Builds on its own prior answers across a session
24	Few-Shot	Pulls a real write-up where someone exploited the same service/version
14	Citation-Aware	Every fact carries its source path — essential for bug-bounty reports

---

🏆 Model benchmark (16 local models, real pentest scenario)

We benchmarked every model in LM Studio on a 2-turn pentest tool-use scenario (recon → exploit chain), measuring speed, accuracy, and valid tool-loops — ejecting each model between runs for clean numbers.

Rank	Model	tok/s	Tool-loops	Verdict
🥇	openai/gpt-oss-20b	165	2/2 ✅	Only model fast, accurate and emitting clean tool calls. 12 GB.
🥈	qwen3.6-35b reasoning-distilled	139	2/2 ✅	Correct chains, needs fence-stripping
—	gemma-4-26b / glm-4.7-flash	156 / 125	0/2 ❌	Fast & accurate, but think without emitting usable tool calls

Key finding: raw "accuracy" is misleading for agents — several high-scoring models produced empty output because they reasoned internally without ever emitting an actionable tool call. The metric that matters is valid tool-loops, and gpt-oss-20b wins it. Reproduce with python model_benchmark.py.

---

🚀 Quick start

Prerequisites

• LM Studio on localhost:1234 (local server enabled) with nomic-embed-text-v1.5 (embeddings) + any chat model loaded — bring your own local model
• Python 3.10+ (tested on 3.14.5)
• A GPU is recommended (built on an RTX 5090; the embedder + a 12 GB chat model fit in 24 GB VRAM — but smaller models work too)

📦 Batteries included: 1,110 curated notes ship in ./vault, so it works the moment you clone. setup.py then pulls in the public doc corpora (HackTricks, PayloadsAllTheThings, OWASP, CTF write-ups) for the full ~60k-chunk brain.

# 1. Clone
git clone https://github.com/test-center-ai/super-rag.git
cd super-rag

# 2. Install (no torch, no Docker, no HuggingFace needed)
pip install -r requirements.txt

# 3. Bootstrap — clone the public doc repos + check LM Studio
python setup.py              # or: python setup.py --minimal  (bundled notes only)

# 4. Build the index (resumable; minutes with parallel embedding)
python main.py ingest

# 5. Ask anything
python main.py query "how do I exploit Apache 2.4.49 path traversal"
python main.py query "AS-REP roasting — what tool and command?"

# 6. Run the agentic pentest loop (hybrid: auto-recon, manual exploit approval)
python main.py pentest 10.10.10.5 --scope 10.10.10.0/24

# 7. Generate a report, or check health
python main.py report
python main.py status

Use your own notes instead? Point it anywhere: SUPERRAG_VAULT=/path/to/your/notes python main.py ingest. Any folder of Markdown works.

Pick a model: set CHAT_MODEL in config.py. Our 16-model benchmark crowned openai/gpt-oss-20b for agentic tool-use; google/gemma-4-12b-qat is a great lighter pick for plain Q&A.

---

🤝 Use it from your AI agents (MCP)

Super RAG runs as one shared MCP server (rag_mcp.py, HTTP @ 127.0.0.1:8765/mcp) so multiple agents can query it concurrently:

python rag_mcp.py        # or let Startup\SuperRAG-MCP.cmd auto-start it

Tools exposed: cybersec_search(query, phase) · cybersec_answer(question) · cybersec_status()

Register it in any MCP client:

// Claude Desktop / OpenClaw style
"mcp": { "servers": { "cybersec-rag": {
  "url": "http://127.0.0.1:8765/mcp", "transport": "streamable-http"
}}}

# Hermes style (config.yaml)
mcp_servers:
  cybersec-rag: { url: http://127.0.0.1:8765/mcp, enabled: true }

See INTEGRATION.md for the full Hermes + OpenClaw walkthrough.

---

🎬 It works — real agent output

Scenario: black-box target, nmap reveals Apache httpd 2.4.49.

STEP 1  recon       → nmap -sS -A 10.10.10.5          ✓ correct first move (2.9s)
STEP 2  enumerate   → gobuster on :80 (Apache live)    ✓ methodical (3.3s)
STEP 3  exploit     → curl --path-as-is "…/cgi-bin/.%2e/…/bin/bash" -d 'reverse shell'
                      ✓ textbook CVE-2021-41773 mod_cgi RCE (4.9s)
        sources: htb-cpts/initial-access-exploitation.md, oswe/file-inclusion-upload.md, HackTricks

The model followed correct methodology, grounded each step in the vault, and produced a working exploit chain — fully offline.

---

📁 Project layout

super-rag/
├── main.py              # CLI: ingest · query · pentest · report · extract · status
├── config.py            # paths, model IDs, chunking, timeouts, tool registry
├── ingest.py            # vault → chunks → parallel embed → Qdrant  (~14× faster pipeline)
├── rag_engine.py        # the 13 RAG strategies + RRF fusion
├── agent.py             # ReAct+Reflect pentest loop, scope + approval gates
├── detector.py          # honeypot / WAF / firewall detection
├── memory.py            # attack-surface graph, findings, stuck-loop tracking
├── report.py            # pentest report + bug-bounty submission generators
├── llm.py               # one streaming chat helper (works for every model)
├── rag_mcp.py           # MCP server for agent integration
├── model_benchmark.py   # the 16-model benchmark harness
├── extract_training.py  # real write-ups → fine-tuning JSONL
└── tools/               # registry + nmap/gobuster/ffuf parsers

---

🧰 Tech stack

Python 3.14 · Qdrant (embedded) · LM Studio (OpenAI-compatible local API) · nomic-embed-text-v1.5 · gpt-oss-20b · MCP / FastMCP — zero cloud dependencies.

---

🔐 Responsible use

• Authorized targets only. Super RAG does not enforce authorization — that is your legal responsibility. Use it on systems you own or have explicit written permission to test (engagements, CTFs, labs).
• Hybrid mode gates exploitation behind a manual [y/N] approval. Don't bypass it.
• Indexed external repos are reference-only and untrusted — never execute code pulled from them.
• This project is for defensive learning, authorized testing, and CTF/education. Don't be a criminal.

---

🛣️ Roadmap

• Qdrant server mode (Docker) for fully-concurrent multi-agent access
• GRPO fine-tune of a 7–14B specialist on extracted real write-ups
• Web UI (the CLI works today)
• Auto-scoping from engagement rules-of-engagement files

🤝 Contributing

Issues and PRs welcome — new tool parsers, RAG strategies, and detector signatures especially. Keep it defensive, keep it cited.

📄 License

MIT — see LICENSE.