Mavent MultiRAG

Advanced RAG pipeline untuk analisis dokumen kompleks.

---

Arsitektur Sistem

graph TB
    subgraph Input
        A[PDF / TXT / DOCX]
    end

    subgraph "Document Pipeline"
        B[Parser]
        C[Structure Detector]
        D[Metadata Extractor]
        E[Semantic Chunker]
        F[Embeddings<br/>nemotron-3-embed-1b]
        G[Chroma Vector Store]
    end

    subgraph "Query Pipeline"
        H[User Query]
        I[Query Rewriter]
        J{Hybrid Retrieval}
        K[Semantic Search<br/>Vector / Chroma]
        L[Keyword Search<br/>BM25]
        M[Reciprocal Rank<br/>Fusion]
        N[Cross-Encoder<br/>Reranker]
        O[Context Selection]
        P[LLM<br/>NVIDIA NIM]
    end

    A --> B --> C --> D --> E --> F --> G
    H --> I --> J
    J --> K
    J --> L
    K --> M
    L --> M
    M --> N --> O --> P

    style A fill:#1a1a2e,color:#fff,stroke:#e94560
    style P fill:#76b900,color:#000,stroke:#fff
    style J fill:#16213e,color:#fff,stroke:#e94560
    style G fill:#0f3460,color:#fff,stroke:#533483

Loading

RAG Pipeline Detail

flowchart LR
    subgraph "Stage 1: Ingest"
        A1[Load] --> A2[Parse]
        A2 --> A3[Structure]
        A3 --> A4[Metadata]
        A4 --> A5[Chunk]
        A5 --> A6[Embed]
    end

    subgraph "Stage 2: Retrieve"
        B1[Rewrite] --> B2{Hybrid}
        B2 --> B3[Vector]
        B2 --> B4[BM25]
        B3 --> B5[Fusion]
        B4 --> B5
        B5 --> B6[Rerank]
    end

    subgraph "Stage 3: Generate"
        C1[Context] --> C2[Prompt]
        C2 --> C3[LLM]
        C3 --> C4[Answer]
    end

    A6 -.-> B1
    B6 -.-> C1

    style A6 fill:#533483,color:#fff
    style B6 fill:#e94560,color:#fff
    style C3 fill:#76b900,color:#000

Loading

Tech Stack

graph LR
    A[Streamlit] --> B[LangChain Classic]
    B --> C[NVIDIA NIM API]
    B --> D[ChromaDB]
    B --> E[rank_bm25]
    B --> F[sentence-transformers]
    C --> G[Nemotron 3 Ultra 550B]
    C --> H[nemotron-3-embed-1b]

    style A fill:#FF4B4B,color:#fff
    style C fill:#76b900,color:#000
    style G fill:#0f3460,color:#fff
    style H fill:#0f3460,color:#fff

Loading

Fitur

Fitur	Deskripsi
Hybrid Retrieval	Semantic (vector) + Keyword (BM25) dengan Reciprocal Rank Fusion
Cross-Encoder Reranking	Precisi ranking pakai ms-marco-MiniLM
Query Rewriting	Pronoun resolution + query expansion otomatis
Document Parsing	Structure-aware: detect headings, tables, sections
Semantic Chunking	Respect sentence/paragraph boundaries
Multi-Format	PDF, TXT, DOCX
7 Model NIM	Nemotron 3 Ultra, Lightning 3.5, Muse Glimmer, dll
Export	Chat, ringkasan, dan insights ke Markdown

Update Terbaru — Hierarchical Async RAG (P95 <5 Menit)

Target: 4–15 PDF × 20 hal → target 2–3 menit, maksimal 5 menit, RPM NIM 40.

Area	Sebelum	Sesudah
Chunking	Fixed 1000/200 RecursiveCharacterTextSplitter	Structure-aware: table → section → para → heading (chunker.py:28), merge <250 chr, is_table/section_title metadata
Summary	Serial 9 batch BATCH_SIZE=15, create_stuff_documents_chain bug Prompt must accept context	Hierarchical map-reduce: chunk (fast, 5/batch, max_tokens 300) → doc → cross-doc synthesis (hierarchical_summarizer.py:30), BATCH_SIZE=40 adaptif, llm.invoke langsung, ETA elapsed/done*remaining
Model Routing	1 model untuk semua	Fast lightning-30b (default, bisa ganti diffusiongemma-26b-a4b-it di Ringkasan) untuk chunk/doc; Reasoning inkling/ultra-550b hanya di sintesis akhir
Pipeline	parse → chunk → embed serial	Parallel: ThreadPool 4 parse, `embed
RAG	invoke blocking	Streaming ChatOpenAI(streaming=True) + pipeline.query_stream + st.write_stream (rag.py:11, chat.py:173)
Chat	Single messages=[], + New Chat hapus histori	Multi-chat conversations={id:[]} + active_conversation_id, PERCAKAPAN jadi tombol klik (sidebar.py:20, chat.py:94)
Charts	2 chart	7 chart: bar kata/halaman/karakter, area proporsi, bar/line chunk, scatter korelasi, bar chunk/doc, histogram (stats.py, insights.py:9)
UI Collapsed	border-right tiang muncul saat hover, + tidak center, menu tidak klik	HAPUS di source app.py:44 border hanya expanded, + centered absolute inset:0 flex, label pointer-events:auto
Cache	vectorstore saja	doc_hash + chunk_hash untuk summary/embed (summary.py: cache, hierarchical_summarizer.py:11)
Benchmark	—	benchmark/fast_model_benchmark.py & stress_15docs_bench.py — Lightning vs DiffusionGemma, concurrency 2,4,8

Cara pakai model cepat: Buka Ringkasan → Model ringkas (cepat) → pilih Lightning 30B (Default) atau DiffusionGemma 26B (Eksperimen) → Generate Ringkasan → progress ETA 2m 30s.

Model yang Tersedia

Model	Type	ID
Nemotron 3 Ultra 550B	Terkuat	nvidia/nemotron-3-ultra-550b-a55b
Nemotron 3.5 Lightning 30B	Seimbang	nvidia/nemotron-3.5-lightning-30b-a3b
Google DiffusionGemma 26B A4B IT	Eksperimen Cepat	google/diffusiongemma-26b-a4b-it
Meta Muse Glimmer 30B	Kreatif	meta/muse-glimmer-30b
StepFun Step 3.7 Flash	Cepat	stepfun-ai/step-3.7-flash
Poolside Laguna XS 2.1	Ringan	poolside/laguna-xs-2.1
Thinking Machines Inkling	—	thinkingmachines/inkling
Embedding: Nemotron 3 Embed 1B	RAG	nvidia/nemotron-3-embed-1b
Fast Summary Default	Cepat	nvidia/nemotron-3.5-lightning-30b-a3b (bisa ganti DiffusionGemma)

Struktur Proyek

mavent-multirag/
├── app.py                         # Entry point (Streamlit)
├── src/smart_doc/
│   ├── config.py                  # Konfigurasi + FAST_SUMMARY_MODELS
│   ├── core/
│   │   ├── document.py            # Document loading
│   │   ├── document_parser.py     # Enhanced parsing + metadata
│   │   ├── chunker.py             # Structure-aware chunking (heading/table/section)
│   │   ├── embeddings.py          # NVIDIA NIM embeddings + async hash cache
│   │   ├── hierarchical_summarizer.py # Chunk→Doc→Cross-doc (fast vs reasoning)
│   │   ├── rag.py                 # RAG chain + LLM (streaming)
│   │   ├── retriever.py           # Hybrid: vector + BM25
│   │   ├── reranker.py            # Cross-encoder reranking
│   │   ├── query_rewriter.py      # Query expansion
│   │   ├── pipeline.py            # Full pipeline orchestrator (parallel parse)
│   │   └── prompts.py             # System prompts
│   ├── ui/
│   │   ├── sidebar.py             # Sidebar + multi-chat
│   │   ├── chat.py                # Chat tab (streaming)
│   │   ├── summary.py             # Summary tab (hierarchical, ETA)
│   │   ├── insights.py            # Key insights tab (7 charts)
│   │   └── stats.py               # Document stats (5+ charts)
│   └── utils/
│       ├── file.py                # File helpers
│       └── export.py              # Export to Markdown
├── benchmark/                     # Benchmark fast models (Lightning vs DiffusionGemma)
│   ├── fast_model_benchmark.py
│   └── stress_15docs_bench.py
├── tests/                         # 20 unit tests + manual
│   ├── test_real_hier_manual.py
│   └── test_real_rag_manual.py
├── plan.md                        # Rencana hierarki P95 <5m
├── .streamlit/config.toml
├── Dockerfile
├── pyproject.toml
└── requirements.txt

Instalasi

git clone https://github.com/Maventlabs/mavent-multirag.git
cd mavent-multirag
python -m venv venv
source venv/bin/activate
pip install -r requirements.txt

Menjalankan

streamlit run app.py

Buka http://localhost:8501 lalu:

1. Masukkan NVIDIA NIM API Key (gratis di build.nvidia.com)
2. Upload dokumen (PDF/TXT/DOCX)
3. Pilih model → mulai tanya

Deploy

Streamlit Community Cloud (gratis)

1. Buka share.streamlit.io
2. Login GitHub → New app → Pilih repo → Deploy

Docker

docker build -t smart-doc-insights .
docker run -p 8501:8501 smart-doc-insights

Testing

python -m unittest discover tests -v

Catatan Keamanan

• API key tidak disimpan permanen
• Dokumen diproses di memori (in-memory)
• Semua model NIM gratis tanpa kartu kredit
• Rate limit: 40 requests/menit

---

_{Dibangun dengan NVIDIA NIM + LangChain + Streamlit}