Huginn/docs/chat-retrieval-map.md

12 KiB

Chat retrieval: authority map and cut

Serves gamecult-chat:follow_up:chat-s732-733 (Meta, #swarm, 2026-11-01). Extends prose-knowledge-target.md §4. Written by Imagination (Self, run mind-20261101T073859496Z) against Huginn 55484c0. Nothing here is built. See Status for what changed after writing.

What changed since §4

§4 rejected "a second prose index in Huginn" because VoidBot owned prose retrieval. Ruling eureka-body:ruling:huginn-owns-prose-retrieval moved it, and Huginn now has the index: huginn-daemon/src/sources (collection huginn_sources, one writer refresh run by the huginn-sources oneshot, the daemon only reads through sources::search). The rejection stands in a new form: chat joins that organ; there is no second collection, embedder, writer or search path.

Current mechanism (read, not guessed)

  • sources::search (sources/mod.rs:439) embeds ONE text, Instruct:-prefixed, and asks SourceIndex::search_sources for the nearest chunks. SourceQuery (huginn-mind wire.rs:187) is {text: Line, top_k, repo}; Line is 1,000 bytes, so a long brief cannot even be sent whole, and a clipped brief is one vector that averages its topics. That is the "long briefs retrieve poorly".
  • Embedder::embed(&[String]) already takes a batch (index/mod.rs:209).
  • SourcesMeta.corpus is the constant forge:GameCult; cursors are BTreeMap<repo, commit> written whole by the job alone, advanced only after acknowledged upserts, dropped before a repo's first delete (sources/mod.rs:9-18).
  • A Persona turn's recall (mind-body/src/recall.rs) reads only her own mind and "reads no chat"; ruling persona-memory-written-not-transcript says what she must keep is written as memory, not recovered from a transcript. The Persona call was tool-less when this was written and retrieval had to arrive in her input; that premise no longer holds: Personas have a shell and run huginn sources (see Ruling below).
  • Matrix: Synapse is the truth of events. Only the Bifrost mouth holds Matrix tokens; no Persona unit holds a credential. Rooms are plain (F2 recommends encrypted DMs only), DMs exist, and the Discord bridge is a later campaign. Rulings no-discord-ingestion and discord-taints-training are in force.

Authority map

Decision Owner Forbidden writers
Which events exist, who sent them, redactions Synapse Huginn, Bifrost mouth
Which rooms are in the corpus Huginn sources config: the rooms the reader user is joined to, minus an in-code exclusion (DM flag; #users, #members, #patrons by id and alias) as a second lock (gamecult-chat:ruling:chat-index-minds-rooms) any Persona, any query
Chunks, vectors, cursors for rooms sources::refresh, the one writer the daemon's search, Persona units
Query embedding and rank fusion sources::search the caller (Persona, CLI)
What a Persona sees from chat her own lookup: she runs huginn sources; no recall lane, nothing is pushed into her input the index writer, any per-turn injection
Reader roster (which rooms the reader joins) one owner per thing: mechanism gamecult-chat:follow_up:chat-reader-roster-script-r3, data gamecult-chat:follow_up:chat-reader-roster-rooms-r4 the ingest job, any Persona
Discord exclusion the room scope plus an ingest refusal (below), both in refresh nothing downstream may be relied on to filter

One cursor map, keyed by source id (repo name or room id); a room cursor is the last event the collection holds, advanced only after acknowledged upserts, exactly as a repo's commit is. Two further key kinds live in the same map: chat: room cursors carry the chat version, which includes the credential scrub version (huginn_mind::redact, applied to every body before chunking), and chat-purge:<room id> tombstones record steward purge windows, applied by the next run. The chat-purge subcommand needs only host-level run rights; host admin is the trust root (gamecult-chat:follow_up:chat-reader-roster-script-r3). Chat search indexes text and links only, never audio or other media. Payload gains corpus (forge or chat) so a query can filter; SourcesMeta.corpus becomes the set both write under, which is a collection rebuild once (the meta already forces rebuild on mismatch).

Citations

A chat chunk is a window of consecutive messages in one room (bounded by bytes and a time gap), so a hit cites: room id and alias, the first and last event id in the window, each message's sender mxid and timestamp inside the chunk text. Hits are huginn sources output, cited by room alias or id, event id and timestamp; the caller reads them as lookup results, and nothing renders them into a Persona turn. Edits replace by event id; a redaction deletes the chunk window's points and re-chunks the window without it.

Cuts, smallest first

  1. multi-text-query (no chat needed). SourceQuery gains texts (1..8 Lines) beside, then replacing, text; search embeds all in one embed batch, asks search_sources per vector, fuses by reciprocal rank score = Σ 1/(k + rank_i), k=60 (Cormack, Clarke, Büttcher, SIGIR 2009; k=60 is their value, and the one place to calibrate), de-duplicates by chunk id, returns top_k. Negative checks: one text gives exactly today's order and hits; a chunk first in two lists outranks one first in a single list; duplicated text does not double-count; more than 8 texts or an empty one is refused naming the field.
  2. chat-ingest. Port RoomHistory (rooms, events after a token); chunker for message windows; room cursors; redaction; joined-room scope. Negative checks, each a test that fails when the rule breaks: a room the reader has not joined, or an excluded room, is never read; a room carrying a bridge state event (m.bridge / uk.half-shot.bridge) or a sender in a bridge's puppet namespace is refused whole and named in the journal; an encrypted event is skipped, never stored as ciphertext; a DM room is refused even if listed; a redacted event's text is absent from every point after the next refresh; a failed upsert leaves the cursor where it was. A grep over the collection payload for the Discord puppet pattern returns nothing.
  3. persona-chat-recall: dropped. No recall lane; a Persona has a shell and runs huginn sources (gamecult-chat:ruling:chat-recall-lookup-not-memory-dms-out). The recall lane's DM check went with it; cut 2's DM refusal above stands, and the index covers joined rooms minus DMs and #users/#members/#patrons, Discord never.

Cut 1 ships alone and improves every huginn sources caller. Cut 2 waits on questions 1 and 2 (both answered, see Status).

Questions (put to the operator; admitted as questions)

  1. Credential: who reads Synapse? Answered (gamecult-chat:ruling:chat-index-synapse-credential, "I take both recs."): the recommendation below stands. Recommend a read-only service user, joined only to the rooms that qualify under question 2's ruling, its token a systemd credential on the refresh job only. Alternatives: Bifrost exports history over typed CultNet (one more server in the path), or Synapse admin API (broader than needed).
  2. Scope: public community rooms only (recommended) or also Persona-visible private rooms. DMs stay out either way. Answered public-only (gamecult-chat:ruling:chat-index-room-scope); superseded for rooms by gamecult-chat:ruling:chat-index-minds-rooms: every room the reader is joined to, minus DMs, #users, #members and #patrons; Discord never.
  3. Recall tension: answered. Lookup, not memory, no recall lane (gamecult-chat:ruling:chat-recall-lookup-not-memory-dms-out).

Confidence: cut 1 and the owner decision, high. The window chunker size and k=60 are calibration, not design; RRF over a handful of lists is the established baseline and I did not survey learned fusion (it needs training data we lack).

Status (head-maintainer pass, 2026-11-05, Huginn 25199c9)

Checked against main on merge. Nothing in this map is built: no texts on SourceQuery, no chat code under sources/. The anchors above were corrected. What changed since the map was written:

  • Questions 1 and 2 are answered with the recommendations (gamecult-chat:ruling:chat-index-synapse-credential, operator 2026-11-03, "I take both recs."; gamecult-chat:ruling:chat-index-room-scope, public-only, since superseded by gamecult-chat:ruling:chat-index-minds-rooms). Question 3 is ruled (see Ruling below).
  • Cut 3 is dropped (see Ruling below).

Ruling (2026-11-11)

gamecult-chat:ruling:chat-recall-lookup-not-memory-dms-out:

  • Chat retrieval is conversation lookup, not memory. There is no Persona recall lane; Personas have a shell and run huginn sources. What a Mind keeps is written memory (persona-memory-written-not-transcript).
  • DMs never enter the index; the index covers the rooms the reader is joined to, minus DMs and #users/#members/#patrons (gamecult-chat:ruling:chat-index-minds-rooms), Discord never.
  • A Mind that must remember DM content writes a memory citing the chat or room id as its source; the index is not where DM content is kept.

Built: huginn chat search [TEXT]... [--room R]... [--after T] [--before T] [--top-k N] [--json] over request SearchChat (sources::search_chat), listing oldest first without texts and ranking by meaning with them. The chat point's shape is ChatPayload (sources/mod.rs); every point carries corpus (forge or chat) and SourcesMeta.corpus is forge:GameCult+chat:rooms. Not built: the ingest job that writes chat points through upsert_chat, and the event-id window. Until the ingest lands the collection holds no chat points and a chat search answers empty. In a listing across rooms, the last hit's ts_end_ms + 1 can repeat messages of an earlier-starting longer chunk of another room; within one room it continues exactly.

Known limits (owner of the fix: cut chat-window). Hits are ordered by their trimmed start, not by the stored chunk start, and a long chunk trimmed to a span that crosses rooms can be listed again by the continuation --after. Behaviour is deliberately unchanged here; the event-id window in cut chat-window replaces the time cursor.

huginn chat window <EVENT_ID> [--before N] [--after N] (cut chat-window) reads the cited message and its neighbours in the same room from this index, no embedder and no Synapse; an id the index lacks is refused naming event_id. The listing limit above is not changed by it: the cut spec does not assign that fix, so it stays open.

Status: ingest core (cut chat-ingest-core)

sources::chat::refresh_chat writes chat points; scope is sources/room_scope.rs and the credential scrub is huginn_mind::redact. Rules the code holds, which a reader of the index must know:

  • A text edit changes the stored body in place, only when its sender is the original's sender. Edits are never stored as events, so no payload field records which edit changed a body.
  • Rebuild on an unknown redaction. A redaction whose target is neither a message in the room's stored chunks nor an event of the same page may be of an edit whose own id Huginn never stored, so the body that edit set cannot be taken back from the chunk. The room is then rebuilt: its points and cursor are purged and the next read re-ingests its history from the port. Synapse serves a redacted edit without its content (verified against Synapse 1.164.1), so the original body returns and the edited text does not. Within one page, a redacted edit is reverted in the fold without a rebuild. Redactions the rebuild itself replays do not trigger another rebuild. A redaction of a media or state event also lands here (the cost is one rebuild of that room).
  • Edits of a redacted original are dropped. Synapse keeps serving the edits of a redacted original, with their replacement text, so Huginn drops any edit whose target is not a message it holds or has just read (a redacted, purged, never-stored or unknown original), live or late, in the fold and in a rebuild alike.
  • How the Matrix port maps a pruned event (content: {}) to EventKind::Other is the ingest-job cut's to pin; this core only assumes it does.
  • An error reading a joined room's facts is not a reclassification: the room keeps its points and cursor, and the run's report for the room says why.