v1.10.0 — VisualWeaver

v1.10.0 — VisualWeaver

RediRecall is now VisualWeaver. The old name carried part of the Redis trademark,
which Redis Ltd.'s policy does not permit, and the Settings panel carried an imitation
of the Redis logo, which it permits still less. Both are gone, and a non-endorsement
notice is in their place.

Alongside the rename, the semantic cache stops refusing to cache whole classes of
question, and the retrieval settings now advise you from your own retrieval history
rather than from a table of recommended values.

Upgrading

This release moves your data directory and the container image. Neither is done for
you — take both steps before starting the new version.

1. Move your data. Everything you own lives here: your configuration and API keys,
your ingestion logs, your uploads, and the Redis data directory holding every chunk you
have ingested. Starting VisualWeaver without moving it looks exactly like total data
loss, because the new build looks in a new place.

# macOS
mv ~/Library/"Application Support"/RediRecall ~/Library/"Application Support"/VisualWeaver

# Linux
mv "${XDG_DATA_HOME:-$HOME/.local/share}"/redirecall "${XDG_DATA_HOME:-$HOME/.local/share}"/visualweaver

Then fix the three absolute paths inside redis/redis.conf, which still point at the
old directory and will stop Redis starting:

sed -i '' 's|/RediRecall/|/VisualWeaver/|g' ~/Library/"Application Support"/VisualWeaver/redis/redis.conf   # macOS
sed -i    's|/redirecall/|/visualweaver/|g' "${XDG_DATA_HOME:-$HOME/.local/share}"/visualweaver/redis/redis.conf  # Linux

2. Pull from the new image path. ghcr.io/sfcyris/redirecall is not updated any
further:

ghcr.io/sfcyris/visualweaver:1.10.0

Environment variables are renamed from REDIRECALL_* to VISUALWEAVER_*. Nothing in
Redis is rewritten and no re-ingest is needed.

Semantic cache

• Asking to visualise something is cached again. It was excluded outright, so asking
  the same question twice cost two model calls — and because nothing was ever stored, a
  later plain-text rewording of it could not match either. Those turns are cached now and
  matched on the question itself rather than on similarity: two different drawing
  requests score closer together than any threshold that would still match anything, so
  plot f(x) = sin(x) must never be answered with sin(2x).
• Ordinary questions are no longer excluded for mentioning a picture. "What is graph
  theory", "explain knowledge graphs in Redis" and "how do I read a flame graph" were all
  treated as requests to draw something and paid again for answers already produced.
• The similarity threshold follows the embedding model. The models do not share a
  scale — two unrelated questions score about 0.10 apart under all-MiniLM-L6-v2 and
  about 0.80 apart under multilingual-e5-base — so a single fixed number was permissive
  under one and admitted unrelated questions under another.
• Threshold and TTL changes take effect without a restart. They were read once when
  the cache was first built, so changing them in Settings did nothing.
• Changing the embedding model clears the cache and its threshold. The stored vectors
  come from the previous model and are not comparable.
• An answer says whether it came from the cache, and if not, why — nothing similar
  stored, an attachment on the turn, or a deliberate regenerate.

Retrieval settings

• Each corpus states the threshold it needs, and the panel reports the value all of
  them can meet. When two genuinely disagree, that disagreement is the finding, with both
  named.
• Scores are kept separately per retrieval mode. Question-mode and HyDE-mode searches
  score differently, and pooling them produced a threshold that suited neither.
• The score distribution is drawn under the slider, with a marker that follows it and
  a live reading of what that value keeps.
• Copy diagnosis puts every measurement behind the verdicts on the clipboard as text.
• The panel says how much traffic its numbers cover when questions were answered from
  cache, since those run no retrieval.
• A setting with nothing measured behind it gets no colour and no suggested value.
• Retrieval statistics survive a restart, and deleting a corpus removes its statistics.