LennyWisdomDB

A SQL Server 2025 vector search sample database. 21K chunks from 270 Lenny's Podcast episodes, pre-embedded and ready to query.

For: SQL Server developers exploring vector search, embeddings, or RAG with real data instead of toy examples.

Quickstart

1. Download LennyWisdomDB.bak (166MB)
2. Restore in SQL Server 2025 (RTM, not preview)
3. Run a semantic search:

DECLARE @query VECTOR(1024);

SELECT @query = search_vector
FROM dbo.search_phrases
WHERE search_phrase = 'what bad strategy looks like and how to avoid it';

SELECT TOP 5
    e.guest_name,
    ec.speaker_question,
    ec.speaker_answer,
    CAST(VECTOR_DISTANCE('cosine', ce.embedding, @query) AS DECIMAL(5,4)) AS distance
FROM dbo.ChunkEmbeddings ce
JOIN dbo.EpisodeChunks ec ON ce.chunk_id = ec.chunk_id
JOIN dbo.Episodes e ON ec.episode_id = e.episode_id
WHERE ec.split_part = 1
ORDER BY VECTOR_DISTANCE('cosine', ce.embedding, @query);

What's Inside

Metric	Value
Episodes	270
Q&A Chunks	21,571
Embedding Model	snowflake-arctic-embed2 (1024 dim)
Pre-embedded Search Phrases	1,076
Topics	87

Features

Feature	How It Works
Vector Search	VECTOR(1024) type + VECTOR_DISTANCE('cosine', ...)
Hybrid Search	Full-text (BM25) + vector with Reciprocal Rank Fusion
Chunk Deduplication	qa_group_id links split chunks to parent Q&A
Contextualized Embeddings	Episode/guest context prepended before embedding

Schema

Content

Table	Purpose
Episodes	Podcast metadata (guest, title, video_id, publish_date)
EpisodeChunks	Q&A segments with timestamps for YouTube deep-linking
Topics / EpisodeTopics	87 topic categories for filtering

Vectors

Table	Purpose
ChunkEmbeddings	1024-dim vector per chunk
EmbeddingModels	Model versioning (name, dimensions, provider)
search_phrases	1,076 pre-embedded queries by category

Deduplication columns in EpisodeChunks:

• qa_group_id - links split chunks to same Q&A
• split_part - 1, 2, 3... for multi-part answers
• content_type - 'intro' or 'main' for pre-filtering
• contextualized_text - enriched text used for embedding

Chunking Strategy

Each chunk = one complete Q&A exchange (Lenny's question + guest's answer). This preserves conversational context naturally.

• Target size: 2,500 chars (~625 tokens)
• Long answers split with 15% overlap, linked via qa_group_id
• Sponsor segments stripped
• Timestamps preserved for YouTube linking

Files

File	Description
LennyWisdomDB.bak	Database backup (166MB)
LennyWisdomDB_Queries.sql	Sample queries (semantic, hybrid, filtered, parent retrieval)

Requirements

• SQL Server 2025 RTM with VECTOR support

Credits

• Transcripts: Lenny Rachitsky via ChatPRD/lennys-podcast-transcripts
• Embeddings: Generated via Ollama (snowflake-arctic-embed2)

License

Schema and scripts: MIT. Transcript content subject to Lenny's original terms.