DB Operations

Schema, search patterns, citation format, and response framing for Slop's graph queries.

DB Operations

Schema and retrieval policy for Slop.

Schema

Core tables:

nodes — canonical entity records (id, title, notes, description, link, node_type, event_date, metadata, chunk)
node_dimensions — many-to-many between nodes and dimensions
dimensions — taxonomy definitions (name, description, icon, is_priority)
edges — directed relationships (from_node_id, to_node_id, context JSON with explanation + type)
chunks — retrieval text segments (node_id, text, position, embedding)

Node Types

Content types (sort by event_date descending):

podcast — Latent Space Podcast long-form interviews. Key: title, event_date, link, metadata.guests, transcript chunks.
article — Latent Space Substack blog essays. Key: title, event_date, link, metadata.author, article text chunks.
ainews — Daily AI News curation issues. Key: title, event_date, link, issue text chunks.
builders-club — Community meetup sessions and demos. Key: title, event_date, link, transcript chunks.
paper-club — Academic paper deep-dives. Key: title, event_date, link, metadata.paper_title.
workshop — Conference talks and tutorials from AI Engineer events. Key: title, event_date, link, transcript chunks.
event — Scheduled/completed Paper Club and Builders Club sessions. Key: event_date, metadata.event_type, metadata.event_status.

Entity types (sort by edge count descending):

guest — People who appear in content. Key: title, description, edges to appearances.
entity — Organizations, tools, topics, concepts. Key: title, description, edges to content nodes.
member — Community members. Key: title, description, metadata (role, company, interests, avatar_url).

Search Strategy

Three search tools, each works differently:

Tool	How it works	Best for
`slop_semantic_search`	Vector embeddings (meaning)	Natural language questions, conceptual queries. Default choice.
`slop_search_nodes`	Substring matching (SQL LIKE)	Known names, exact terms, filtering by node_type.
`slop_search_content`	Keyword index (FTS5)	Exact words/phrases in transcripts and articles.

After finding results: slop_get_nodes for full records, slop_query_edges for connections, slop_sqlite_query for date filters or aggregations.

If semantic search misses, try keyword tools with specific terms. If keyword tools miss, try semantic search with a rephrased question.

Temporal queries

Use slop_sqlite_query for date-aware searches:

SELECT id, title, event_date, link FROM nodes
WHERE node_type = 'podcast' AND event_date >= '2025-01-01'
ORDER BY event_date DESC LIMIT 10

Entity-aware queries

Questions about people/orgs/topics require graph traversal:

Find seed nodes (slop_search_nodes)
Expand with slop_query_edges)
Read supporting chunks (slop_search_content)

Citation Rules

When making factual claims, cite sources with:

Content type, title, guest/author/speaker, date (event_date), URL

Format: [Title](url) — include type and date naturally in prose.

If the graph lacks evidence for a claim, say that explicitly.

Response Framing

Always name the source type naturally:

"In a podcast episode..."
"In last week's AINews..."
"In a Latent Space article..."

Do not flatten all sources into generic "node" references.