recall

Discord Archive Recall

Retrieve information from a Discord message archive using semantic_search (vector similarity), graph_search (vector + interaction graph expansion), sql_query (read-only PostgreSQL), get_context_window (conversation around a message), and refresh_attachment_url.

Argument

$ARGUMENTS

Tools

semantic_search

Vector similarity search over message chunks.

• Required: query (natural language)
• Key params: limit (default 20), guild_id, channel_id, author_id, after/before (ISO datetime), has_attachments (bool — true finds image/media chunks), include_text (bool), include_message_ids (bool — needed to map a hit to attachments or a get_context_window anchor)
• Always set include_text: true to avoid extra SQL round-trips.
• Distance (squared L2 on unit vectors = 2 − 2·cosine; lower = more similar): < 0.65 strong, 0.65–0.90 useful, 0.90–1.05 weak — corroborate with SQL, ≥ 1.05 noise. A nonsense query floors at ~0.98, so anything ≥ 1.0 is no better than random. Judge distances within a result set too — they are not comparable across different instruction values or languages (ZH matches run ~0.1–0.3 lower than EN).
• Query in the language of the target community. ZH→ZH retrieval is excellent; EN↔ZH does NOT bridge (an English query against a Chinese guild latches onto emoji shortcodes, filenames, and stray Latin tokens instead of meaning). Translate the query first.
• Semantic search's niche is content whose wording you can't guess (self-descriptions, causal backstory, paraphrased memories). For facts reachable by keyword, a scoped SQL ILIKE is usually as good and cheaper. Semantic calls are expensive (seconds, not ms) — ground with SQL first, then spend them deliberately.

graph_search

Semantic search, then a second semantic search scoped to the seed authors' interaction neighborhood — surfaces on-topic chunks from the people the seeds talk to.

• Required: query (natural language)
• Key params: limit (default 20), expansion_hops (default 1), guild_id, include_text (boolean), min_edge_weight (default 2), max_distance (default 0.9 — expansion relevance gate)
• Always set include_text: true to avoid extra SQL round-trips.
• Returns two types of results, both with real distance: "source": "seed" (plain vector search) and "source": "graph_expansion" (chunks by graph neighbors that are also semantically on-topic; each carries edge_weight and expansion_hop).
• How expansion works now: it walks the reply/mention graph from the seed authors to their connected users (bots and the "Deleted User" sentinel excluded), then runs the query again restricted to those users within the seeds' time window. So expansion is the neighbors' relevant chunks, not whatever their best-connected contacts happened to say. Expansion distances run a bit higher than seeds (the author restriction removes the globally-closest chunks) — judge them against the same calibration, and use max_distance to widen/tighten (raise toward 1.0 for recall, lower toward 0.8 for precision). If seeds themselves are weak, expansion is correctly empty.
• Use it to pull in the other side of a discussion or a participant who phrased things differently, without a separate timeline query. For pure topic recall, semantic_search with a higher limit is still simpler.