Skip to main contentThe shared agents form the foundational layer of the haus²⁵ curation system, providing essential services that all specialized agents depend on. These agents implement advanced LangChainJS patterns and cost optimization strategies to ensure efficient operation across the entire multi-agent system.
Agent Architecture
RAG Agent
The RAG Agent provides contextual user history and preferences using LangChain’s vector storage capabilities with significant cost optimizations.
Technical Implementation
Core Architecture: GoogleGenerativeAIEmbeddings with text-embedding-004 model, MemoryVectorStore for in-memory operations, and file-based document persistence.
User History Indexing
Blockchain Data Integration: Fetches creator events from EventFactory contract, retrieves metadata from IPFS via Pinata gateway, creates Document objects with event content and metadata, adds to vector store with deduplication checking.
Cost Optimization Features
Document Deduplication:
- Checks existing cache before adding new documents
- Prevents redundant embeddings generation
- Reduces API calls to Google Embeddings
File-Based Persistence:
- Avoids expensive vector database costs
- Maintains state across service restarts
- Enables quick startup with cached embeddings
Research Agent
The Research Agent provides market intelligence and trend analysis using multiple data sources with intelligent caching and cost reduction through LangChain summarization.
Multi-Source Data Collection
Google Custom Search Integration: Category-specific trend research with 6-month date restriction, intelligent caching to avoid redundant API calls.
YouTube Data API Integration: Recent video analysis for performance trends and audience insights.
Intelligent Analysis: Gemini-powered insights generation with length constraints and JSON structure for pricing, timing, audience, trends, and keywords optimized for RTA events.
Memory Agent
The Memory Agent implements the on-chain iteration system, providing persistent, cost-effective AI memory storage.
On-Chain Storage Pattern
EventManager Integration: Planner proxy private key with global whitelist permissions, SEI testnet wallet client with public actions extension for contract interactions.
Advantages Over Traditional Storage
Cost Comparison:
- Traditional vector DB: $300-840/year
- On-chain storage: sub-$10/year for 1000 events
- Cost reduction: 100x savings
Consistency Benefits:
- Single source of truth on blockchain
- No cache invalidation issues
- Persistent across all deployments
- Universal access from any client
Blockchain Agent
The Blockchain Agent manages SEI testnet integration and proxy delegation patterns for secure contract interactions.
Proxy Management
Scope-Specific Initialization: Dynamic private key selection based on curation scope (planner, promoter, producer) with fallback to planner proxy for unknown scopes.
Trends Agent
The Trends Agent provides advanced social media trend analysis using Apify for data collection and LangChain summarization chains for cost optimization.
Apify Integration: Dynamic client initialization with token authentication, cached results with timestamp tracking for efficiency.
LangChain Summarization
Token Reduction Strategy: Document creation from Twitter content, TokenTextSplitter with 3000 chunk size and 200 overlap, LoadSummarizationChain with “refine” type for cost-effective processing of large social media datasets.
Social Knowledge Agent
The Social Knowledge Agent maintains up-to-date platform specifications and requirements for multi-platform content optimization.
Comprehensive Platform Support: Text limits, image specifications, video requirements, and optimal posting strategies for X/Twitter, Facebook, Instagram, and EventBrite platforms.
Integration Patterns
Shared Context Preparation
Coordinated Data Collection: Parallel execution of user history indexing, category research, and trends analysis to avoid redundant API calls across supervisors.
Cost Tracking: Token estimation across RAG embeddings (500), research analysis (1500), trends analysis (2000), and memory operations (100) for total ~4100 tokens per event.
