querying-from-seekdb

name: querying-from-seekdb description: "Query and export data from seekdb vector database. Supports two search modes: (1) Scalar search - metadata filtering only, (2) Hybrid search - fulltext + semantic search combined. The --query-text parameter is used for BOTH fulltext ($contains) and semantic (query_texts) search simultaneously. Can export results to CSV/Excel." license: MIT

Query and Export Data from seekdb

Query data from seekdb vector database with support for scalar search, hybrid search (fulltext + semantic), and export to CSV/Excel files.

Path Convention

Note: All paths in this document (e.g., scripts/) are relative to THIS skill directory, not the project root.

Prerequisites

• Python 3.10+ installed
• Data imported into seekdb collection
• Required packages:

pip install pyseekdb pandas openpyxl

⚠️ CRITICAL: Execution Workflow

MUST FOLLOW this workflow when handling user search requests:

Step 1: Get Collection Information (If Not Already Known)

Before constructing any query, you MUST understand the data structure. However, you should cache this information within the conversation.

Caching Rules:

• ✅ First query for a collection: Execute --info to get metadata structure
• ✅ Subsequent queries for the SAME collection: Use cached info from earlier in conversation, skip --info
• ✅ Query for a DIFFERENT collection: Execute --info for the new collection
• ✅ User explicitly asks for collection info: Execute --info

# Get collection info to see metadata fields (only if not already known)
python scripts/query_from_seekdb.py <collection_name> --info

This shows:

• Total record count
• Available metadata field names (e.g., source, year, category)
• Sample documents

Example conversation flow:

User: "找 seekdb_demo 中 2023 年的教程"
→ Claude Code: 执行 --info (第一次查询此 collection)
→ 发现 metadata 有 source, year 字段
→ 执行搜索

User: "再找一下 notion 来源的"
→ Claude Code: 不需要再执行 --info (同一 collection，结构已知)
→ 直接执行搜索

User: "查一下 another_collection 中的数据"
→ Claude Code: 执行 --info (不同 collection)
→ 了解新 collection 的结构
→ 执行搜索

Step 2: Analyze User Request

Parse the user's natural language request to identify:

Important: --query-text is used for BOTH fulltext search ($contains) and semantic search (query_texts) simultaneously. The same text is used for both.

Step 3: Choose Search Method

User Request Analysis
       │
       ▼
┌─────────────────────────────────────────────────────────────┐
│ Does the request involve ONLY metadata field conditions?    │
│ (e.g., "year=2023", "source=notion", no content search)     │
└─────────────────────────────────────────────────────────────┘
       │
       ├── YES ──► Scalar Search: --where only
       │
       └── NO ───► Does it involve content/semantic search?
                          │
                          ├── YES (no metadata) ──► Hybrid Search: --query-text only
                          │
                          └── YES (with metadata) ──► Scalar + Hybrid: --where + --query-text

Two Search Modes

Mode 1: Scalar Search (Metadata Only)

When to use: User wants to filter by metadata fields ONLY, no content/semantic search needed.

# Filter by metadata fields only
python scripts/query_from_seekdb.py seekdb_demo --where '{"source": "notion", "year": 2023}'

Example requests:

• "找出所有来自 notion 的文档"
• "显示 2023 年的记录"
• "source 是 google-docs 的数据"

Mode 2: Hybrid Search (Fulltext + Semantic)

When to use: User wants to search by content - the query text is used for BOTH fulltext matching AND semantic similarity.

# Hybrid search: query text used for both fulltext ($contains) and semantic (query_texts)
python scripts/query_from_seekdb.py seekdb_demo --query-text "seekdb 教程"

How it works:

• --query-text "seekdb 教程" → Fulltext: where_document: {"$contains": "seekdb 教程"} + Semantic: query_texts: "seekdb 教程"
• Results are ranked using RRF (Reciprocal Rank Fusion)

Example requests:

• "找 seekdb 教程" → --query-text "seekdb 教程"
• "搜索 python 技术文档" → --query-text "python 技术文档"

Mode 3: Scalar + Hybrid Search

When to use: User wants metadata filtering + content/semantic search.

# Metadata filter + Hybrid search
python scripts/query_from_seekdb.py seekdb_demo --query-text "seekdb 教程" --where '{"year": 2023}'

Example requests:

• "请找出 seekdb_demo 集合中 2023 年写的 seekdb 教程" → --query-text "seekdb 教程" --where '{"year": 2023}'
• "找 notion 来源的编程指南" → --query-text "编程指南" --where '{"source": "notion"}'

🎯 Real-World Example Analysis

User request: "请找出 seekdb_demo 集合中 2023 年写的 seekdb 教程"

Step 1: Run --info to get metadata structure:

python scripts/query_from_seekdb.py seekdb_demo --info
# Output shows metadata fields: source, year

Step 2: Analyze request:

Step 3: Execute:

python scripts/query_from_seekdb.py seekdb_demo --query-text "seekdb 教程" --where '{"year": 2023}'

CLI Reference

Commands

# List all collections
python scripts/query_from_seekdb.py --list-collections

# Show collection info (run this first to understand data structure!)
python scripts/query_from_seekdb.py <collection_name> --info

# Scalar search (metadata filter only)
python scripts/query_from_seekdb.py <collection_name> --where '<json_filter>'

# Hybrid search (fulltext + semantic, using same query text for both)
python scripts/query_from_seekdb.py <collection_name> --query-text "<text>" [-n <count>]

# Scalar + Hybrid search (metadata filter + fulltext + semantic)
python scripts/query_from_seekdb.py <collection_name> --query-text "<text>" --where '<json>'

# Export to CSV/Excel
python scripts/query_from_seekdb.py <collection_name> <search_options> --output results.csv
python scripts/query_from_seekdb.py <collection_name> <search_options> --output results.xlsx

Options

Filter Operators

How to Construct --where Parameter

Step 1: Run --info to see available metadata fields:

python scripts/query_from_seekdb.py seekdb_demo --info

# Example output:
# Collection: seekdb_demo
#   Total records: 2
# Preview (first 3 records):
#   ID: doc1...
#     Document: python tutorial...
#     Metadata keys: ['source', 'year']    ← These are the metadata field names!

Step 2: Use the metadata field names to construct --where:

# From the output above, we know the collection has 'source' and 'year' fields
# So we can filter by these fields:

--where '{"source": "notion"}'           # source equals "notion"
--where '{"year": 2023}'                 # year equals 2023
--where '{"source": "notion", "year": 2023}'  # both conditions (implicit AND)

Step 3: Match user request to metadata fields:

Metadata Filter Operators

Complex Filter Examples

# Multiple conditions with implicit AND (both must be true)
--where '{"source": "notion", "year": 2023}'

# Explicit AND
--where '{"$and": [{"source": "notion"}, {"year": {"$gte": 2023}}]}'

# OR condition
--where '{"$or": [{"source": "notion"}, {"source": "google-docs"}]}'

# Range condition (year between 2022 and 2024)
--where '{"$and": [{"year": {"$gte": 2022}}, {"year": {"$lte": 2024}}]}'

# Combined AND + OR
--where '{"$and": [{"year": 2023}, {"$or": [{"source": "notion"}, {"source": "obsidian"}]}]}'

Export to CSV/Excel

# Export scalar search results to CSV
python scripts/query_from_seekdb.py mobiles --where '{"Brand": "SAMSUNG"}' --output samsung.csv

# Export hybrid search results to Excel
python scripts/query_from_seekdb.py mobiles --query-text "good camera" --output results.xlsx

# Export with custom sheet name
python scripts/query_from_seekdb.py mobiles --query-text "phone" --output phones.xlsx --sheet-name "Search Results"

Supported Export Formats

Data Structure in seekdb

seekdb stores data in two distinct locations:

Connection Configuration

Set environment variables for server mode:

References

• pyseekdb SDK Getting Started
• hybrid_search API
• get API
• Filter Operators