moss-semantic（苔藓语义）

Moss Agent Skills

Capabilities

Moss is the real-time semantic search runtime for conversational AI. It delivers sub-10ms lookups and instant index updates that run in the browser, on-device, or in the cloud - wherever your agent lives. Agents can create indexes, embed documents, perform semantic/hybrid searches, and manage document lifecycles without managing infrastructure. The platform handles embedding generation, index persistence, and optional cloud sync - allowing agents to focus on retrieval logic rather than infrastructure.

Skills

Index Management

• Create Index: Build a new semantic index with documents and embedding model selection
• Load Index: Load an existing index from persistent storage for querying
• Get Index: Retrieve metadata about a specific index (document count, model, etc.)
• List Indexes: Enumerate all indexes under a project
• Delete Index: Remove an index and all associated data

Document Operations

• Add Documents: Insert or upsert documents into an existing index with optional metadata
• Get Documents: Retrieve stored documents by ID or fetch all documents
• Delete Documents: Remove specific documents from an index by their IDs

Search & Retrieval

• Semantic Search: Query using natural language with vector similarity matching
• Keyword Search: Use BM25-based keyword matching for exact term lookups
• Hybrid Search: Blend semantic and keyword search with configurable alpha weighting
• Metadata Filtering: Constrain results by document metadata (category, language, tags)
• Top-K Results: Return configurable number of best-matching documents with scores

Embedding Models

• moss-minilm: Fast, lightweight model optimized for edge/offline use (default)
• moss-mediumlm: Higher accuracy model with reasonable performance for precision-critical use cases

SDK Methods

API Actions

All REST API operations go through POST /manage with an action field:

• createIndex - Create index with seed documents
• getIndex - Get metadata for single index
• listIndexes - List all project indexes
• deleteIndex - Remove index and assets
• addDocs - Upsert documents into index
• getDocs - Retrieve stored documents
• deleteDocs - Remove documents by ID

Workflows

Basic Semantic Search Workflow

1. Initialize MossClient with project credentials
2. Call createIndex() with documents and model (moss-minilm or moss-mediumlm)
3. Call loadIndex() to prepare index for queries
4. Call query() with search text and top_k parameter
5. Process returned documents with scores

Hybrid Search Workflow

1. Create and load index as above
2. Call query() with alpha parameter to blend semantic and keyword
3. alpha: 1.0 = pure semantic, alpha: 0.0 = pure keyword, alpha: 0.6 = 60/40 blend
4. Default is semantic-heavy (\~0.8) for conversational use cases

Document Update Workflow

1. Initialize client and ensure index exists
2. Call addDocs() with new documents and upsert: true option
3. Existing documents with matching IDs are updated; new IDs are inserted
4. Call deleteDocs() to remove outdated documents by ID

Voice Agent Context Injection Workflow

1. Initialize MossClient and load index at agent startup
2. On each user message, automatically query Moss for relevant context
3. Inject search results into LLM context before generating response
4. Respond with knowledge-grounded answer (no tool-calling latency)

Offline-First Search Workflow

1. Create index with documents using local embedding model
2. Load index from local storage
3. Query runs entirely on-device with sub-10ms latency
4. Optionally sync to cloud for backup and sharing

Integration

Voice Agent Frameworks

• LiveKit: Context injection into voice agent pipeline with inferedge-moss SDK
• Pipecat: Pipeline processor via pipecat-moss package that auto-injects retrieval results

Context

Authentication

SDK requires project credentials:

• MOSS_PROJECT_ID: Project identifier from Moss Portal
• MOSS_PROJECT_KEY: Project access key from Moss Portal

export MOSS_PROJECT_ID=your_project_id
export MOSS_PROJECT_KEY=your_project_key

REST API requires headers:

• x-project-key: Project access key
• x-service-version: v1: API version header
• projectId in JSON body

Package Installation

Document Schema

interface DocumentInfo {
  id: string; // Required: unique identifier
  text: string; // Required: content to embed and search
  metadata?: object; // Optional: key-value pairs for filtering
}

Query Parameters

Model Selection

Performance Expectations

• Sub-10ms local queries (hardware-dependent)
• Instant index updates without reindexing entire corpus
• Sync is optional; compute stays on-device
• No infrastructure to manage

Chunking Best Practices

• Aim for \~200–500 tokens per chunk
• Overlap 10–20% to preserve context
• Normalize whitespace and strip boilerplate

Common Errors

Async Pattern

All SDK methods are async - always use await:

// JavaScript
await client.createIndex("faqs", docs, "moss-minilm");
await client.loadIndex("faqs");
const results = await client.query("faqs", "search text", 5);

# Python
await client.create_index("faqs", docs, "moss-minilm")
await client.load_index("faqs")
results = await client.query("faqs", "search text", top_k=5)

For additional documentation and navigation, see: https://docs.usemoss.dev/llms.txt