tools-only Architecture Overview
OpenViking is a context database designed for AI Agents, unifying all context types (Memory, Resource, Skill) into a directory structure with semantic retrieval and progressive content loading.
System Overview
┌────────────────────────────────────────────────────────────────────────────┐
│ OpenViking System Architecture │
├────────────────────────────────────────────────────────────────────────────┤
│ │
│ ┌─────────────┐ │
│ │ Client │ │
│ │ (OpenViking)│ │
│ └──────┬──────┘ │
│ │ │
│ ┌─────────────────────────┼─────────────────────────┐ │
│ │ │ │ │
│ ▼ ▼ ▼ │
│ ┌─────────────┐ ┌─────────────┐ ┌─────────────┐ │
│ │ Retrieve │ │ Session │ │ Parse │ │
│ │ (Context │ │ (Session │ │ (Context │ │
│ │ Retrieval) │ │ Management)│ │ Extraction)│ │
│ │ search/find │ │ add/used │ │ Doc parsing │ │
│ │ Intent │ │ compress │ │ L0/L1/L2 │ │
│ │ Rerank │ │ commit │ │ Tree build │ │
│ └──────┬──────┘ └──────┬──────┘ └──────┬──────┘ │
│ │ │ │ │
│ │ │ Memory extraction │ │
│ │ ▼ │ │
│ │ ┌─────────────┐ │ │
│ │ │ Compressor │ │ │
│ │ │ Compress/ │ │ │
│ │ │ Deduplicate │ │ │
│ │ └──────┬──────┘ │ │
│ │ │ │ │
│ └────────────────────────┼────────────────────────┘ │
│ ▼ │
│ ┌─────────────────────────────────────────────────────────────────┐ │
│ │ Storage Layer │ │
│ │ AGFS (File Content) + Vector Index │ │
│ └─────────────────────────────────────────────────────────────────┘ │
│ │
└────────────────────────────────────────────────────────────────────────────┘
Core Modules
| Module | Responsibility | Key Capabilities |
|---|---|---|
| Client | Unified entry | Provides all operation interfaces, coordinates modules |
| Retrieve | Context retrieval | Intent analysis (IntentAnalyzer), hierarchical retrieval (HierarchicalRetriever), Rerank |
| Session | Session management | Message recording, usage tracking, session compression, memory commit |
| Parse | Context extraction | Document parsing (PDF/MD/HTML), tree building (TreeBuilder), async semantic generation |
| Compressor | Memory compression | 6-category memory extraction, LLM deduplication decisions |
| Storage | Storage layer | VikingFS virtual filesystem, vector index, AGFS integration |
Dual-Layer Storage
OpenViking uses a dual-layer storage architecture separating content from index (see Storage Architecture):
| Layer | Responsibility | Content |
|---|---|---|
| AGFS | Content storage | L0/L1/L2 full content, multimedia files, relations |
| Vector Index | Index storage | URIs, vectors, metadata (no file content) |
Data Flow Overview
Adding Context
Input → Parser → TreeBuilder → AGFS → SemanticQueue → Vector Index
- Parser: Parse documents, create file and directory structure (no LLM calls)
- TreeBuilder: Move temp directory to AGFS, enqueue for semantic processing
- SemanticQueue: Async bottom-up L0/L1 generation
- Vector Index: Build index for semantic search
Retrieving Context
Query → Intent Analysis → Hierarchical Retrieval → Rerank → Results
- Intent Analysis: Analyze query intent, generate 0-5 typed queries
- Hierarchical Retrieval: Directory-level recursive search using priority queue
- Rerank: Scalar filtering + model reranking
- Results: Return contexts sorted by relevance
Session Commit
Messages → Compress → Archive → Memory Extraction → Storage
- Messages: Accumulate conversation messages and usage records
- Compress: Keep recent N rounds, archive older messages
- Archive: Generate L0/L1 for history segments
- Memory Extraction: Extract 6-category memories from messages
- Storage: Write to AGFS + vector index
Deployment Modes
Embedded Mode
For local development and single-process applications:
client = OpenViking(path="./data")
- Auto-starts AGFS subprocess
- Uses local vector index
- Singleton pattern
Service Mode
For production distributed deployment:
client = OpenViking(
vectordb_url="http://vectordb:5000",
agfs_url="http://agfs:8080"
)
- Connects to remote services
- Supports multiple concurrent instances
- Independently scalable
Design Principles
| Principle | Description |
|---|---|
| Pure Storage Layer | Storage only handles AGFS operations and basic vector search; Rerank is in retrieval layer |
| Three-Layer Information | L0/L1/L2 enables progressive detail loading, saving token consumption |
| Two-Stage Retrieval | Vector search recalls candidates + Rerank improves accuracy |
| Single Data Source | All content read from AGFS; vector index only stores references |
Related Documents
- Context Types - Resource/Memory/Skill types
- Context Layers - L0/L1/L2 model
- Viking URI - Unified resource identifier
- Storage Architecture - Dual-layer storage details
- Retrieval Mechanism - Retrieval process details
- Context Extraction - Parsing and extraction process
- Session Management - Session and memory management