lancer（lancer 文档）

Lancer - LanceDB CLI and MCP Server Skill

You are a specialist in using lancer, a CLI and MCP server for LanceDB that provides semantic and full-text search with multi-modal support (text and images). This skill provides comprehensive workflows, best practices, and common patterns for document ingestion, search, and table management.

What is Lancer?

lancer is a powerful tool for:

• Semantic search: Find documents by meaning, not just keywords
• Multi-modal support: Index and search both text and images
• LanceDB integration: Efficient vector database storage and retrieval
• Flexible ingestion: Support for multiple file formats (txt, md, pdf, sql, images)
• MCP server mode: Integration with Claude Desktop and other MCP clients

Core Capabilities

1. Ingest: Add documents to LanceDB with automatic chunking and embedding
2. Search: Semantic similarity search across documents
3. Tables: Manage LanceDB tables (list, info, delete)
4. Remove: Remove documents from tables
5. MCP: Run as Model Context Protocol server

Quick Start

Basic Search

# Search all tables
lancer search "how to deploy kubernetes"

# Search specific table with more results
lancer search -t docs -l 20 "authentication methods"

# Search with similarity threshold
lancer search --threshold 0.7 "error handling patterns"

Basic Ingestion

# Ingest a single file
lancer ingest document.md

# Ingest a directory
lancer ingest ./docs/

# Ingest multiple paths
lancer ingest file1.md file2.pdf ./images/

Document Ingestion

Ingest Command Options

# Ingest to specific table
lancer ingest -t my_docs document.md

# Ingest with file extension filter
lancer ingest -e md,txt,pdf ./docs/

# Ingest from stdin (pipe file paths)
find ./docs -name "*.md" | lancer ingest --stdin

# Ingest from file list
lancer ingest --files-from paths.txt

# Custom chunk size and overlap
lancer ingest --chunk-size 2000 --chunk-overlap 400 document.md

Supported File Types

Text formats:

• txt - Plain text files
• md - Markdown documents
• pdf - PDF documents
• sql - SQL scripts

Image formats:

• jpg, jpeg - JPEG images
• png - PNG images
• gif - GIF images
• bmp - Bitmap images
• webp - WebP images
• tiff, tif - TIFF images
• svg - SVG vector graphics
• ico - Icon files

Embedding Models

Text models:

# Default: all-MiniLM-L6-v2 (fast, good quality)
lancer ingest document.md

# Larger model for better quality
lancer ingest --text-model all-MiniLM-L12-v2 document.md

# BGE models (better semantic understanding)
lancer ingest --text-model bge-small-en-v1.5 document.md
lancer ingest --text-model bge-base-en-v1.5 document.md

Image models:

# Default: clip-vit-b-32 (cross-modal text/image)
lancer ingest image.jpg

# ResNet50 for image-only search
lancer ingest --image-model resnet50 image.jpg

Advanced: Force specific model:

# Force CLIP for text (enables future image additions)
lancer ingest --embedding-model clip-vit-b-32 document.md

# Force BGE for performance (text-only)
lancer ingest --embedding-model BAAI/bge-small-en-v1.5 document.md

Ingestion Optimization

# Filter by file size
lancer ingest --min-file-size 1000 --max-file-size 10000000 ./docs/

# Skip embedding generation (metadata only)
lancer ingest --no-embeddings document.md

# Custom batch size for database writes
lancer ingest --batch-size 200 ./large-dataset/

# JSON output for scripting
lancer ingest --format json document.md

Search Operations

Search Command Options

# Basic search
lancer search "kubernetes deployment"

# Search specific table
lancer search -t docs "authentication"

# Limit results
lancer search -l 5 "error handling"

# Set similarity threshold (0.0-1.0)
lancer search --threshold 0.6 "database migration"

# Include embeddings in results
lancer search --include-embeddings "API design"

# JSON output
lancer search --format json "machine learning"

Metadata Filters

# Single filter (field:operator:value)
lancer search --filter "author:eq:John" "AI research"

# Multiple filters
lancer search \
  --filter "author:eq:John" \
  --filter "year:gt:2020" \
  "deep learning"

# Available operators:
# eq (equals), ne (not equals)
# gt (greater than), lt (less than)
# gte (greater/equal), lte (less/equal)
# in (in list), contains (string contains)

Search Examples

# Find recent documentation
lancer search \
  -t docs \
  --filter "date:gte:2024-01-01" \
  -l 10 \
  "API endpoints"

# Search by category
lancer search \
  --filter "category:eq:tutorial" \
  "getting started"

# Multi-criteria search
lancer search \
  -t technical_docs \
  --filter "language:eq:python" \
  --filter "level:eq:advanced" \
  --threshold 0.7 \
  -l 15 \
  "async programming patterns"

Table Management

List Tables

# List all tables
lancer tables list

# JSON output
lancer tables list --format json

Table Information

# Get table details
lancer tables info my_table

# JSON output for scripting
lancer tables info my_table --format json

Delete Table

# Delete a table (be careful!)
lancer tables delete old_table

Remove Documents

# Remove specific documents from a table
lancer remove -t docs document_id

# Remove multiple documents
lancer remove -t docs id1 id2 id3

Configuration

Using Config File

# Specify config file
lancer -c ~/.lancer/config.toml search "query"

# Set default table in config
lancer -c config.toml ingest document.md

Environment Variables

# Set default table
export LANCER_TABLE=my_docs
lancer search "query"  # Searches my_docs

# Set log level
export LANCER_LOG_LEVEL=debug
lancer ingest document.md

Log Levels

# Error only
lancer --log-level error search "query"

# Warning
lancer --log-level warn ingest document.md

# Info (default)
lancer --log-level info search "query"

# Debug
lancer --log-level debug ingest document.md

# Trace (verbose)
lancer --log-level trace search "query"

Common Workflows

Workflow 1: Index Documentation

# 1. Ingest markdown docs
lancer ingest -t docs -e md ./documentation/

# 2. Verify ingestion
lancer tables info docs

# 3. Test search
lancer search -t docs "installation guide"

# 4. Refine search with threshold
lancer search -t docs --threshold 0.7 -l 5 "configuration"

Workflow 2: Multi-modal Image Search

# 1. Ingest images with CLIP model
lancer ingest -t images -e jpg,png,webp \
  --image-model clip-vit-b-32 \
  ./photos/

# 2. Search images with text query
lancer search -t images "sunset over mountains"

# 3. Search with higher threshold for precision
lancer search -t images --threshold 0.8 "red car"

Workflow 3: Mixed Content Corpus

# 1. Ingest with CLIP for cross-modal search
lancer ingest -t knowledge_base \
  --embedding-model clip-vit-b-32 \
  -e md,pdf,jpg,png \
  ./content/

# 2. Search text and images together
lancer search -t knowledge_base "architecture diagrams"

# 3. Filter by file type
lancer search -t knowledge_base \
  --filter "file_type:eq:png" \
  "system design"

Workflow 4: Batch Ingestion

# 1. Generate file list
find ./corpus -type f -name "*.md" > files.txt

# 2. Ingest from list with custom settings
lancer ingest -t corpus \
  --files-from files.txt \
  --chunk-size 1500 \
  --chunk-overlap 300 \
  --batch-size 150

# 3. Verify ingestion
lancer tables info corpus

# 4. Test search quality
lancer search -t corpus -l 10 "sample query"

Workflow 5: Update Existing Corpus

# 1. Ingest new documents
lancer ingest -t docs ./new_docs/

# 2. Search to verify new content
lancer search -t docs "recent feature"

# 3. Remove outdated documents
lancer remove -t docs old_doc_id

# 4. Verify final state
lancer tables info docs

Best Practices

1. Choose the Right Embedding Model

For text-only corpora:

# Fast and efficient
lancer ingest --text-model all-MiniLM-L6-v2 document.md

# Better quality
lancer ingest --text-model bge-base-en-v1.5 document.md

For images or mixed content:

# Cross-modal search (text queries → image results)
lancer ingest --embedding-model clip-vit-b-32 content/

2. Optimize Chunk Settings

Short documents (< 500 words):

lancer ingest --chunk-size 500 --chunk-overlap 100 article.md

Long documents (> 2000 words):

lancer ingest --chunk-size 2000 --chunk-overlap 400 book.pdf

Code documentation:

lancer ingest --chunk-size 1000 --chunk-overlap 200 docs/

3. Use Tables to Organize Content

# Separate tables by content type
lancer ingest -t api_docs ./api/*.md
lancer ingest -t tutorials ./tutorials/*.md
lancer ingest -t images ./screenshots/*.png

# Search specific context
lancer search -t api_docs "authentication endpoints"

4. Set Appropriate Thresholds

Broad exploration:

lancer search --threshold 0.4 "general topic"

Precise matching:

lancer search --threshold 0.75 "specific concept"

Very high precision:

lancer search --threshold 0.85 -l 3 "exact information"

5. Use Filters for Structured Data

# Combine semantic search with metadata
lancer search \
  --filter "status:eq:published" \
  --filter "category:eq:tutorial" \
  --threshold 0.6 \
  "getting started guide"

6. Format Output for Scripting

# JSON output for automation
lancer search --format json "query" | jq '.results[] | .path'

# List tables programmatically
lancer tables list --format json | jq '.[] | .name'

MCP Server Mode

Running as MCP Server

# Start MCP server for Claude Desktop integration
lancer mcp

# With custom config
lancer mcp -c ~/.lancer/config.toml

# With specific log level
lancer mcp --log-level info

Integration with Claude Desktop

Add to Claude Desktop config (~/Library/Application Support/Claude/claude_desktop_config.json):

{
  "mcpServers": {
    "lancer": {
      "command": "lancer",
      "args": ["mcp"]
    }
  }
}

Performance Tips

1. Batch Operations

# Ingest multiple files at once
lancer ingest file1.md file2.md file3.md

# Use --stdin for large batches
find ./docs -name "*.md" | lancer ingest --stdin

2. Optimize Batch Size

# Larger batches for bulk ingestion
lancer ingest --batch-size 500 ./large-corpus/

# Smaller batches for limited memory
lancer ingest --batch-size 50 ./documents/

3. Skip Embeddings for Metadata-Only

# Index metadata without generating embeddings
lancer ingest --no-embeddings ./archive/

4. Use Appropriate Models

# Faster ingestion with smaller model
lancer ingest --text-model all-MiniLM-L6-v2 ./docs/

# Better quality with larger model (slower)
lancer ingest --text-model bge-base-en-v1.5 ./docs/

Troubleshooting

Issue: Search returns no results

Solutions:

# Lower the similarity threshold
lancer search --threshold 0.3 "query"

# Check table exists and has documents
lancer tables list
lancer tables info my_table

# Try different search terms
lancer search "alternative phrasing"

Issue: Ingestion fails for some files

Solutions:

# Check supported extensions
lancer ingest -e md,txt,pdf ./docs/

# Set file size limits
lancer ingest --max-file-size 100000000 ./docs/

# Use debug logging
lancer --log-level debug ingest document.pdf

Issue: Low search quality

Solutions:

# Use better embedding model
lancer ingest --text-model bge-base-en-v1.5 document.md

# Adjust chunk size
lancer ingest --chunk-size 1500 --chunk-overlap 300 document.md

# Adjust search threshold
lancer search --threshold 0.6 "query"

Issue: Slow ingestion

Solutions:

# Increase batch size
lancer ingest --batch-size 300 ./docs/

# Use faster embedding model
lancer ingest --text-model all-MiniLM-L6-v2 ./docs/

# Skip embeddings if not needed
lancer ingest --no-embeddings ./docs/

Quick Reference

# Ingestion
lancer ingest document.md                          # Ingest single file
lancer ingest -t docs ./directory/                 # Ingest to specific table
lancer ingest -e md,pdf ./docs/                    # Filter by extensions
lancer ingest --chunk-size 2000 document.md        # Custom chunk size

# Search
lancer search "query"                              # Search all tables
lancer search -t docs "query"                      # Search specific table
lancer search -l 20 "query"                        # Limit results
lancer search --threshold 0.7 "query"              # Set similarity threshold
lancer search --filter "author:eq:John" "query"    # Metadata filter

# Table management
lancer tables list                                 # List all tables
lancer tables info my_table                        # Table information
lancer tables delete old_table                     # Delete table

# Configuration
lancer -c config.toml search "query"               # Use config file
lancer --log-level debug ingest doc.md             # Set log level
export LANCER_TABLE=docs                           # Set default table

# MCP server
lancer mcp                                         # Start MCP server

Common Patterns

Pattern 1: Quick Documentation Search

lancer search -t docs --threshold 0.7 -l 5 "how to configure authentication"

Pattern 2: Ingest and Test

lancer ingest -t test_docs document.md && \
lancer search -t test_docs "key concept from document"

Pattern 3: Find Similar Images

lancer search -t images --threshold 0.8 "sunset landscape photography"

Pattern 4: Batch Ingest with Verification

find ./docs -name "*.md" | lancer ingest -t docs --stdin && \
lancer tables info docs

Pattern 5: Precise Technical Search

lancer search -t technical_docs \
  --filter "language:eq:rust" \
  --threshold 0.75 \
  -l 10 \
  "async trait implementation patterns"

Summary

Primary use cases:

• Semantic search across documentation
• Multi-modal search (text and images)
• Knowledge base indexing and retrieval
• Integration with Claude via MCP

Key advantages:

• Semantic similarity (not just keyword matching)
• Multi-modal support (text and images)
• Flexible metadata filtering
• Multiple embedding model options
• Fast vector search with LanceDB

Most common commands:

• lancer ingest document.md - Index documents
• lancer search "query" - Search semantically
• lancer tables list - Manage tables
• lancer search -t docs --threshold 0.7 "query" - Precise search