Mirror notice (29 Sep 2026). This repository is a mirror. The canonical SIDIX archive is fahmiwol/sidix. This mirror now carries the same reviewed content: unreviewed working notes were taken down, as they were in the canonical repository on 16 Sep 2026. The full historical source is at tag
v0.7.0of the canonical repository.
Free & Open Source AI Agent
Self-Hosted · Self-Learning · Self-Evolving · Own Stack · No Vendor API
🔬 MiganCore research method · Metode riset MiganCore · 📖 Quran Lab · 🧾 Authorship · Kepenemuan
🌐 Website · 🧠 The Foundation · 🏗️ Architecture · 🤝 Contribute
EN — SIDIX was an open research build from April to August 2026. The hosted app was discontinued on purpose and will not return. Anyone who wants to use SIDIX will download the code and run it on their own machine; the useful parts of the code are being repackaged in a readable, documented form. Until that package is published, this repository holds this description, the license and authorship statement, and the Quran Lab provider. Everything further down this page is a snapshot of that build.
The research continues in MiganCore — an own, Indonesian-first language model and the measurement discipline built around it. SIDIX's sanad (chain-of-transmission) idea lives on there: six of its trained models so far carry a provenance chain that is checked link by link.
ID — SIDIX adalah pembangunan riset terbuka dari April sampai Agustus 2026. Aplikasi yang di-hosting sengaja dihentikan dan tidak akan dihidupkan lagi. Yang ingin memakai SIDIX cukup mengunduh kodenya dan menjalankannya di mesin sendiri; bagian kode yang berguna sedang dikemas ulang agar mudah dibaca dan terdokumentasi. Sampai paket itu terbit, repositori ini memuat deskripsi ini, lisensi dan pernyataan kepenemuan, serta provider Quran Lab. Semua di bawah bagian ini adalah potret build tersebut.
Riset berlanjut di MiganCore — model bahasa milik sendiri yang mengutamakan bahasa Indonesia, beserta disiplin pengukuran di sekelilingnya. Gagasan sanad SIDIX hidup terus di sana: sejauh ini enam model yang dilatih membawa rantai asal-usul yang diperiksa mata demi mata.
| Read the continuation · Baca kelanjutannya | English | Bahasa Indonesia |
|---|---|---|
| MiganCore research method · Metode riset MiganCore | README | README |
| From SIDIX to MiganCore: the ecosystem · Ekosistem | 01 | 01 |
| The research method · Metode | 02 | 02 |
| Measurement integrity · Integritas pengukuran | 03 | 03 |
| Hallucination · Halusinasi | 05 | 05 |
| Case studies · Studi kasus | 07 | 07 |
Contact · Kontak: Fahmi Ghani — fahmiwol@gmail.com · collaboration welcome · terbuka untuk kolaborasi
EN: Quran Lab — engineering interpretations with citations and explicit unreviewed status.
ID: Quran Lab — interpretasi rekayasa dengan sitasi dan status belum ditinjau yang dinyatakan jelas.
Not a chatbot. SIDIX is an AI Agent with initiative, opinions, and creativity. It brainstorms with you, builds for you, and grows from every conversation. Self-hosted. MIT licensed. Yours forever.
| Vol | Feature | Impact |
|---|---|---|
| 20-fu3 | Simple-tier fast-path (greetings/ack) | 78s → 2s (37× speedup) |
| 20 | Semantic cache L2 + BGE-M3 embedding | <100ms warm, multilingual ID |
| 20 | Complexity router (simple/standard/deep) |
Auto-route reasoning depth |
| 20 | Domain detector (fiqh/medis/coding/factual) | Per-domain cache threshold + sanad gating |
| 20 | Style anomaly filter (BadStyle defense) | Corpus poisoning prevention |
| 19 | Relevance + Quality Sprint (4 modules) | Better retrieval ranking |
| 17 | CodeAct enrich + MCP wrap | Code blocks auto-execute |
| 16 | Creative Agent Ecosystem (10 domain × 37 agent) | Multi-agent debate/iteration |
| 15 | LoRA SIDIX adapter on Qwen2.5-7B | Self-trained, 4-bit QLoRA |
Production stack (historical, April–August 2026): VPS (FastAPI brain · BGE-M3 CPU · 2.287 corpus docs) + RunPod GPU serverless (vLLM v2.14.0 · Qwen2.5-7B + LoRA).
Direction: AI Agent yang BEBAS dan TUMBUH — bebas dari single-prompt loop, tumbuh compound dari setiap interaksi.
"SIDIX adalah entitas kecerdasan komprehensif yang tidak hanya mengeksekusi perintah multi-modal, tetapi secara PROAKTIF mengevaluasi, memori-optimasi, dan mengorkestrasi ekosistem tools untuk menciptakan nilai komersial dan inovasi TANPA PENGAWASAN TERUS-MENERUS."
- 🧠 The Mind — Self-Correction & Metacognition · Distributed Long-Term Memory (RAG) · Chain/Tree of Thoughts
- ✋ The Hands & Tools — Tool Orchestration & API Mastery · Aesthetic & Commercial Judgement · Resource Management
- 🚀 The Drive — Intrinsic Proactivity (no prompt needed) · Boundary & Context Awareness
- 🧠 Memory — 5-layer immutable + LoRA + RAG (no catastrophic forgetting)
- 🎭 Multi-Agent — Innovator (Burst) + Critic + Tadabbur 3-persona convergence
- 🔄 Continuous Learning — auto_lora + rehearsal buffer + nightly retrain
- 🤖 Proactive — anomaly detect + self-prompt + daily digest (gerak sendiri)
UTZ (creative/visual) · ABOO (engineer/technical) · OOMAR (strategist/business) · ALEY (academic/research) · AYMAN (general/hangat)
👁 melihat · 👂 mendengar · 🗣 berbicara · ✋ merasakan · 🤲 1000 tangan paralel (design + code + riset + posting bersamaan)
"The measure of intelligence is not how much you know, but how precisely you know what you don't know — and how honestly you say so."
Most AI tools are black boxes controlled by corporations. You pay per token. Your data trains their model. You have no idea what they do with it.
SIDIX is built on a different premise:
- Free — run it without paying anyone per-query
- Open source — every line of code is auditable
- Self-hosted — your server, your data, your model
- Self-learning — improves from real usage, structured, every quarter
Inspired by a 1,400-year-old knowledge system, SIDIX asks: what if the architecture of knowledge matters more than its volume?
What if an AI that knows why it knows, how it knows, and the limits of what it knows — is more trustworthy than one that simply knows a lot?
That question is the origin of IHOS — and the reason SIDIX exists.
IHOS (Islamic Holistic Ontological System) is not a religious constraint. It is an epistemological architecture — a framework for how knowledge should be structured, validated, and used.
It was derived from observing one of the oldest self-expanding knowledge systems ever recorded.
Consider this: a single text of ~77,000 words has generated 14 centuries of derivative scholarship — yielding jurisprudence, medicine, cosmology, mathematics, linguistics, ethics, and governance. Different readers, in different times and places, derive entirely different — yet internally consistent — bodies of knowledge from the same source.
This is not a coincidence of literary richness. It is a designed architecture.
The classical Islamic scholars identified it precisely:
| Qur'anic Layer | Technical Analog | SIDIX Implementation |
|---|---|---|
| Zahir — the explicit text | Frozen Foundation Model | Qwen2.5-7B + LoRA — immutable base weights |
| Batin — the latent meaning | Latent space / embeddings | BM25 + vector corpus — contextual retrieval |
| Asbabun Nuzul — grounded context | Grounded Generation | Every output grounded in brand_brief + user_state + platform_context |
| Sanad — the chain of transmission | Provenance tracking | [FACT] / [OPINION] / [UNKNOWN] labels + citation chain |
| Maqashid — the higher objectives | Objective function | 5-axis filter: life · intellect · faith · lineage · wealth |
| Ijtihad — reasoned interpretation | Agentic reasoning | agent_react.py ReAct loop |
| Naskh — abrogation/update | Knowledge conflict resolution | naskh_handler.py — sanad-tier based supersession |
| Tafakkur — deliberate reflection | Meta-cognition | muhasabah_loop.py — Niyah→Amal→Muhasabah self-refinement |
| Tadrij — progressive revelation | Curriculum learning | curriculum_engine.py L0→L4 knowledge ladder |
The key insight: Al-Qur'an's power doesn't come from size — it comes from architecture. A frozen core. Context-sensitive derivation. Verified transmission chains. Purpose-aligned interpretation. This is what SIDIX translates into code.
1. FROZEN CORE, LIVING EDGES
The base model doesn't retrain arbitrarily — it grows through structured LoRA adapters.
Like a text that never changes, but whose understanding deepens with the reader.
2. SOURCE CHAIN IS NON-NEGOTIABLE
Every claim is labeled. Every output has a traceable path.
Not because we're cautious — because honesty is a design constraint.
3. CONTEXT IS FIRST-CLASS INPUT
User state, brand context, time, platform — these are not metadata.
They are part of the inference function.
4. PURPOSE FILTERS KNOWLEDGE
Not all technically correct answers are appropriate answers.
Maqashid as objective function: output is evaluated against human flourishing, not just accuracy.
5. GROWTH IS STRUCTURAL, NOT INCIDENTAL
Self-learning isn't a feature. It's the architecture.
Daily corpus ingestion → curation → LoRA retrain → deploy. Every quarter, SIDIX improves.
SIDIX is not a chatbot with a nice UI. It's a three-layer cognitive agent running 100% on your own stack:
┌─────────────────────────────────────────────────────────────┐
│ LAYER 1 — BRAIN (LLM) │
│ Local LLM + optional adapter │
│ Generative inference — token by token, own stack │
│ No vendor API required for default mode │
└─────────────────────────┬───────────────────────────────────┘
│ ReAct loop (agent_react.py)
┌─────────────────────────▼───────────────────────────────────┐
│ LAYER 2 — HANDS (Tools + RAG) │
│ 35 active tools: │
│ ├── Knowledge: search_corpus · read_chunk · concept_graph │
│ ├── Web: web_fetch · web_search · pdf_extract │
│ ├── Code: code_sandbox · code_analyze · code_validate │
│ ├── Creative: generate_copy · brand_kit · plan_campaign │
│ ├── Image: text_to_image (SDXL self-hosted) │
│ ├── Multi-Agent: Raudah Protocol (parallel specialists) │
│ └── Growth: roadmap_* · workspace_* · muhasabah_refine │
└─────────────────────────┬───────────────────────────────────┘
│ daily cycle
┌─────────────────────────▼───────────────────────────────────┐
│ LAYER 3 — MEMORY (Growth Loop) │
│ 50+ open sources → corpus queue → curation → JSONL │
│ → adapter retrain (offline pipeline) → adapter deploy │
│ SIDIX gets smarter every quarter. Structurally. │
└─────────────────────────────────────────────────────────────┘
New in v0.6: Raudah (روضة المعرفة — Garden of Knowledge) — a multi-agent parallel system where specialists work concurrently and the Orchestrator synthesizes a consensus answer.
Task → RaudahOrchestrator.urai_task()
→ IHOS Guardrail (Maqashid check)
→ asyncio.gather([Researcher, Analyst, Writer, Engineer, Verifier])
→ RaudahOrchestrator.agregasi() ← Ijma' (consensus synthesis)
→ RaudahResult.jawaban_final
Unlike "swarm" architectures: no vendor API, IHOS guardrail before spawn, Sanad Validator per output.
import asyncio
from brain.raudah.core import run_raudah
result = asyncio.run(run_raudah("Research 5 productive waqf models in Southeast Asia"))
print(result.jawaban_final) # synthesized consensus
print(result.durasi_s) # e.g. 45.2s on RTX 3060SIDIX can optionally expose a local plugin server for compatible clients (e.g., Claude Desktop, Cursor, GPT Actions, Codex). This is not required for the default standing-alone mode.
Compatible client
│ MCP (stdio)
▼
apps/sidix-mcp/src/index.js ← 13 tools total
│
├── SIDIX brain backend
├── Extension bridge (optional)
└── Messaging bridge (optional)
Install & run (optional):
See integration docs under docs/ for client-specific setup. This is an adapter path, not a dependency.
Social tools (example):
| Tool | What it does |
|---|---|
scan_instagram_profile |
ER + sentiment + tier dari profil publik IG |
scan_threads_profile |
Analisis profil Threads |
scan_youtube_channel |
Engagement rate YouTube channel |
scan_twitter_profile |
Analisis X/Twitter profil |
analyze_social |
Analisis mendalam dari URL apapun |
compare_social_accounts |
Banding 2+ akun lintas platform |
social_post_threads |
Auto-post ke Threads (butuh token) |
wa_send |
Kirim pesan via messaging bridge (opsional) |
wa_receive |
Baca inbox via messaging bridge (opsional) |
| Domain | Agent / Tool | Status |
|---|---|---|
| Coding | code_sandbox · code_analyze · code_validate · project_map |
✅ Live |
| Self-awareness | self_inspect — SIDIX reads its own tool registry |
✅ Live |
| Copywriting | generate_copy (AIDA/PAS/FAB, 3 variants) |
✅ Live |
| Content Strategy | generate_content_plan (7/14/30-day calendar) |
✅ Live |
| Brand Building | generate_brand_kit (name + archetype + palette + voice) |
✅ Live |
| Visual Content | generate_thumbnail + text_to_image (SDXL) |
✅ Live |
| Campaign | plan_campaign (AARRR funnel + KPI) |
✅ Live |
| Ads | generate_ads (FB/Google/TikTok copy) |
✅ Live |
| Quality Gate | muhasabah_refine (CQF ≥ 7.0 loop) |
✅ Live |
| Multi-Agent | Raudah Protocol v0.1 (parallel specialists, local backbone) | ✅ Live |
| Knowledge Conflict | Naskh Handler (sanad-tier based resolution) | ✅ Live |
| Maqashid Filter | v2 mode-based: CREATIVE/ACADEMIC/IJTIHAD/GENERAL | ✅ Live |
| Self-Evolution | prompt_optimizer — L1 flywheel, weekly improvement |
✅ Live |
| Knowledge | BM25 corpus · Wikipedia · web_search · web_fetch | ✅ Live |
| Image | Image generation (local-first) | ✅ Live |
| Social Intelligence | scan_instagram_profile · scan_threads · scan_youtube · scan_twitter · analyze_social · compare_social_accounts |
✅ Live |
| Messaging Automation | wa_send · wa_receive (optional bridge) |
✅ Live |
| Plugin Server | Optional plugin server (13 tools via stdio MCP) | ✅ Live |
| Chrome Extension | Social Radar MV3 — DOM scrape + background service worker | ✅ Live |
| Voice / Video | Whisper + TTS + FFmpeg | 🗓 Sprint 8 |
| 3D / Gaming | Hunyuan3D + Blender API | 🗓 Sprint 9 |
| Raudah v0.2 | TaskGraph DAG + POST /raudah/run endpoint | 🗓 Next sprint |
SIDIX adapts its voice, depth, and framing based on who it's talking to. Each persona maps to a Maqashid mode — a different lens for evaluating and presenting knowledge.
| Persona | Character | Specialization | Maqashid Mode |
|---|---|---|---|
| AYMAN | Strategic Sage | Research synthesis, long-form, Islamic epistemology, vision | IJTIHAD |
| ABOO | The Analyst | Data, logic, structured argument, code review, decisions | ACADEMIC |
| OOMAR | The Craftsman | Technical deep-dives, system design, build & implementation | IJTIHAD |
| ALEY | The Learner | Teaching, curriculum, beginner-friendly, patient explanation | GENERAL |
| UTZ | The Generalist | Daily tasks, creative work, conversational, quick answers | CREATIVE |
# Auto-routing — SIDIX picks the right persona from your question
from brain_qa.persona import route_persona
result = route_persona("help me design a logo for a tech startup")
# → PersonaDecision(persona='AYMAN', confidence=0.63, reason='signal=creative/design')
# Or specify explicitly
from brain_qa.agent_react import run_react
session = run_react(question="audit this Python function", persona="ABOO")Backward compatible: legacy persona aliases are accepted internally, but never surfaced in public UI/content.
EN — The hosted app is discontinued and the source code is not on this branch. A clean, documented package for running SIDIX on your own machine is being prepared; this section will point to it once it is published.
ID — Aplikasi yang di-hosting sudah dihentikan dan kode sumbernya tidak ada di cabang ini. Paket yang rapi dan terdokumentasi untuk menjalankan SIDIX di mesin sendiri sedang disiapkan; bagian ini akan menautkannya begitu terbit.
SIDIX now has a soul — 7-pillar self-awareness system:
- Nafs (Pilar 1): 7-topic routing with persona character injection
- Aql (Pilar 2): Self-learning — every good interaction becomes training data (CQF ≥7.0)
- Qalb (Pilar 3): Health monitoring — auto-heals on degradation
- Hayat (Pilar 5): Self-iteration — refines answers when quality is low
Optional local plugin server for compatible clients (disabled by default).
Complete 7-pillar standalone modules in brain/ directory:
brain/nafs/— 3-layer knowledge fusion (60% parametric + 30% KG + 10% static)brain/aql/— Jariyah v2: capture→CQF→validate→storebrain/qalb/— SyifaHealer: 4-level health monitoringbrain/hayat/— generate→evaluate→refine loopbrain/ruh/— weekly evaluation + improvement planningbrain/ilm/— knowledge gap detection + auto-crawlbrain/hikmah/— QLoRA retrain trigger
SIDIX understands Indonesian slang, abbreviations, and typos gracefully. 4-layer stack: Normalizer → Semantic Matcher → Confidence Scorer → Context Responder
Host integration bridge (optional).
- 13 tools: 4 core (query, capture, learn, status) + 9 social intelligence
- Chrome Extension + WA Bridge + Extension Bridge
- OpenAPI spec (tooling integration)
- 7-topic routing (ngobrol/umum/kreatif/koding/sidix_internal/agama/etika)
- Persona character injection per response
- Self-learning training pairs (Aql)
- Health monitoring (Qalb)
- Nafs 3-layer wire to main agent
- Jariyah v3 real-time capture (thumbs feedback)
- Branch system (multi-client)
- PostgreSQL schema
- FLUX.1 image generation
- TTS (Piper)
- Code validator
- Sidebar UI Framework
- Image Editor v1
- Branch selector
- Raudah v2 multi-agent
- Brand Guidelines Maker
- Auto-retrain trigger
v0.9 beta target: Q3 2026 — Agency Kit + Raudah v2 + multimodal parity
For those who want to go deeper:
# Sanad (chain of transmission) → citation chain in every output
{
"answer": "...",
"label": "[FACT]", # Zahir — what is explicitly stated
"citations": [ # Sanad — who said it, where
{"source": "...", "sanad_tier": "primer", "chunk_id": "..."}
],
"maqashid_filter": "passed" # Maqashid — does this serve human flourishing?
}
# Naskh (abrogation) → knowledge conflict resolution
from brain_qa.naskh_handler import NaskhHandler
handler = NaskhHandler()
winner, status, reason = handler.resolve(old_item, new_item)
# → ("superseded", "New source has higher sanad tier (primer > aggregator)")
# Tafakkur (deliberate reflection) → muhasabah loop
def muhasabah_loop(output, brief):
niyah = validate_intent(brief) # Was the intention clear?
amal = score_cqf(output) # Was the action good? (CQF ≥ 7.0)
review = reflect_on_gaps(output) # What can be improved?
return refine(output) if amal < 7.0 else output
# Maqashid v2 — mode-based, not keyword blacklist
from brain_qa.maqashid_profiles import evaluate_maqashid, MaqashidMode
result = evaluate_maqashid(
user_query="Write copy for our coffee brand",
generated_output="Bold flavor, honest origin...",
mode=MaqashidMode.CREATIVE # → BOOST intellect & wealth, no block
)
# → {"status": "pass", "tagged_output": "...\n\n[Intellect-Optimized | Value-Creation Mode]"}EN — The most useful contributions now are scholarly review of the Quran Lab studies and discussion of the MiganCore research method. Open an issue in either repository.
ID — Kontribusi yang paling berguna sekarang adalah tinjauan ilmiah atas kajian Quran Lab dan diskusi metode riset MiganCore. Buka issue di salah satu repositori.
- ✅ No vendor API in inference pipeline — zero data leaves your server
- ✅ G1 Safety Policy — anti-injection, anti-PII, anti-toxic
- ✅ Maqashid v2 — intent-based filter, not keyword blacklist (creative-safe)
- ✅ Audit log (append-only, hash-chained) for every tool call
- ✅ Identity masking for public-facing endpoints
- ✅ 4-label epistemic tagging — hallucinations are labeled, not hidden
Building agents that keep their own state? These are the pieces I extracted from doing it:
- Agent Memory Starter — free, open source. Four memory types and three prompts.
- Agent Memory OS — $3. The full twelve-prompt method.
- Second Brain Kit — $7. One memory for every agent, served over MCP.
- MCP Server Starter — $5. Build an MCP server, then prove it behaves.
All of them: fahmiwolf.gumroad.com
MIT License — see LICENSE. The list of original methods and the authorship statement are in AUTHORSHIP.md.
Lisensi MIT — lihat LICENSE. Daftar metode orisinal dan pernyataan kepenemuan ada di AUTHORSHIP.md.
Use it. Fork it. Teach it. Build on it.
Project website: sidixlab.com
"We don't build AI that replaces human judgment. We build AI that makes human judgment more informed."
