Files

T

Gmer4Lfe d9225d06d9 Route a General Chat question to the profile that fits it

2026-08-09 00:36:13 -04:00

16 KiB

Raw Blame History

━━━━━ AI ━━━━━

Retrieval over Varaverk's own documentation. Ask the system a question about itself and get an answer grounded in its actual headers, READMEs, Manuals and conf templates — with sources.

Everything here is off by default and optional. Varaverk works exactly as well with AI_ENABLED=false as with it true. Nothing in the ecosystem depends on this folder.

File	Role
`ai_index.sh`	Build / refresh the retrieval index
`ai_query.sh`	Ask a question, or search the index directly
`lib/chunk.js`	Split repo files into retrieval units
`lib/index.js`	Embed chunks, store vectors in SQLite
`lib/search.js`	Embed a query, score it, rank results
`lib/cli.js`	Argument bridge the bash entry points call

━━━ WHY THIS WORKS AT ALL ━━━

Because the header audit came first.

The single worst failure in naive retrieval is a chunk that contains half of one idea and half of another — a fixed-size window cutting mid-thought, embedding two unrelated things as one vector. That problem does not exist here, because every script in the repo carries the same six sections at exact, greppable boundaries:

PURPOSE → OPERATIONAL MODEL → DESIGN PRINCIPLES → OPERATIONAL SAFEGUARDS
        → CONFIGURATION → RUNTIME MODES

Split on those and every chunk is a coherent unit by construction. No token windows, no overlap heuristics, no tuning.

And because the headers say why. A model can read mover_stop.sh and describe what it does. It cannot look at that code and know the cache writers are lockless on purpose, or that removeCompletedDownloads being true on both arrs is intentional. Those live in DESIGN PRINCIPLES and OPERATIONAL SAFEGUARDS, which is precisely what makes this index worth more than an equivalent pile of source.

━━━ SECTION ROUTING ━━━

Every chunk stores its section name as its own column, so a question's shape can steer retrieval before similarity is even considered:

Question shape	Steered toward
"what stops X and Y overlapping"	`OPERATIONAL SAFEGUARDS`
"which variable controls X"	`CONFIGURATION`
"does this take --dry-run"	`RUNTIME MODES`
"why is it built this way"	`DESIGN PRINCIPLES`
"how does X work"	`OPERATIONAL MODEL`

Applied as a score boost, not a filter. Intent detection is a heuristic, and a heuristic must never be able to exclude the one chunk that holds the answer. --section=NAME forces a hard filter when you actually want one.

The boost still has a blind spot worth knowing about: definitional questions. "What is Varaverk?" matches the PURPOSE intent, so every script's one-line PURPOSE gets boosted above the top-level prose that actually answers it — and the model correctly replies that the context does not define the system. The corpus is fine; README.md is indexed. The routing simply buries it. --kind=readme is the hard filter for that case:

bash AI/ai_query.sh --kind=readme "what is Varaverk"

--kind filters on where a chunk came from — header, readme, manual, template, doc, ui — and composes with --section. Prefer it over --section for "what is" and "why does this exist" questions, where the answer is narrative rather than a header field.

━━━ NAMED-PARAGRAPH SUB-CHUNKING ━━━

Sections alone were not granular enough, and the failure was instructive.

rsync.sh documents fourteen distinct safeguards in one 2.8k-character OPERATIONAL SAFEGUARDS block. Asked "what happens if pass 1 of a merge run fails", the correct chunk scored 0.558 — below unrelated chunks from other files — because the other thirteen safeguards dominated the vector.

The header convention writes each safeguard as a named paragraph: an unindented title, an indented body. Splitting on those titles took the same query to 0.718 and first place.

Merge-Run Delete Interlock            ← its own chunk
  --delete is applied only when pass 1 completed...

The parent section name is carried onto every sub-chunk, so routing still works. The title detection requires the next line to be indented — without that check, any wrapped prose line became a spurious boundary mid-sentence.

━━━ WHAT IS INDEXED ━━━

Roughly 2,900 chunks across ~180 files:

Kind	Source
`header`	Script headers — bash and PHP, sections and named paragraphs
`readme`	Every `README-*.md`
`manual`	Every `Manual-*.md`
`template`	`Deployment/*.template` — the versioned conf schema
`doc`	Top-level `README.md`, `Manual.md`, design notes
`ui`	`Plugin/unraid/pages/readme/*.md` — the WebGUI's own help panels

Script bodies are not indexed. Headers state intent, code states mechanism; for the questions this answers, intent retrieves better and costs far less.

━━━ THE SAFETY BOUNDARY ━━━

Only git ls-files is ever indexed. This is not a convenience — it is the security model.

Configurations/ and data/ are gitignored, so every file holding a credential was never in the repo to begin with. The index therefore describes the full conf schema (via the tracked templates, which carry all the explanatory comments) while structurally never containing a secret.

Do not "improve" this into a filesystem walk. A logged secret can be rotated. A secret averaged into a 768-dimension float cannot be found, let alone removed.

A live conf value the model genuinely needs should arrive through a tool call at query time, subject to redaction — never baked into a vector at index time.

━━━ HOW IT IS BUILT ━━━

SQLite, no vector database. ~2,900 chunks × 768 dims is a couple of million multiply-adds per query — under a millisecond. A vector DB would be a container to run, monitor, fail over and back up, in exchange for nothing at this scale.

Vectors are raw little-endian float32 BLOBs. nomic-embed-text returns L2-normalised vectors, so cosine similarity is a plain dot product — no normalising, no magnitude cache. PHP reads the same blobs with unpack('f*', $blob) when the UI needs them.

Incremental on mtime. A file whose mtime has not moved is skipped without being read. A no-op refresh takes about 70 ms; a full rebuild takes a few minutes.

Node for the maths, bash for everything else. The bash entry points own configuration, gating, locking and logging exactly as every other Varaverk job does. Node owns only float vector maths and SQLite BLOBs. Same split as api_cache_writer.sh and its PHP.

lib/index.js uses node:sqlite, which Node still marks experimental. It is used because it needs no native compilation on Unraid. If a Node upgrade ever breaks it, the index is regenerable in minutes — this is a disposable artefact, not a datastore.

━━━ USAGE ━━━

# Build (needs AI_ENABLED=true)
bash AI/ai_index.sh                  # incremental
bash AI/ai_index.sh --force          # full rebuild
bash AI/ai_index.sh --dry-run        # what would be indexed; contacts nothing
bash AI/ai_index.sh --status         # size, counts, last build

# Ask
bash AI/ai_query.sh "what stops rsync and the mover running at once"
bash AI/ai_query.sh --search "why are the cache writers lockless"
bash AI/ai_query.sh --section=CONFIGURATION "which variable sets the mover grace period"
bash AI/ai_query.sh --kind=readme "what is Varaverk"          # definitional / narrative
bash AI/ai_query.sh --json "..."     # for other scripts

--search is the trustworthy mode. It returns verbatim repo text with nothing generated. When the answer matters, use it — or read the sources the generated answer cites.

━━━ TRUST THE SOURCES, NOT THE PROSE ━━━

The generation prompt instructs the model to answer only from retrieved context and to say what is missing rather than fill the gap from general knowledge. That instruction matters here more than usual: this repo's conventions are frequently not the conventional ones, and a confident generic answer about rsync, Docker or systemd is worse than no answer.

It works — and its faithfulness cut both ways on the first real test. Asked which variable controls the mover's grace window, the model answered MOVER_STOP_TIMEOUT, "defaults to 30 seconds", citing mover_stop.sh › CONFIGURATION. The variable was right. The 30 was wrong — the real value is 300 — and the model was quoting the header verbatim. The header was stale.

That sweep then found six stale (default: N) claims across the repo, all since corrected. The lesson is the operating principle for this whole folder:

Retrieval is exactly as accurate as the documentation it points at. When an answer looks wrong, check the cited source before blaming the model — it is usually reporting a real problem in the repo.

━━━ WHAT THIS DOES NOT DO ━━━

It does not write conf. AI_CONF_WRITE_ENABLED exists in master.conf and is off, with an empty key whitelist. Nothing in this folder writes a setting.
It does not make decisions. No watchdog, cleanup or fallback path consults it. The per-feature AI_ASSIST_* toggles are all false and each one is earned separately.
It does not index code. Headers and docs only.
It does not sync. The index is host-local and gitignored. Each node builds its own.
It is not required. Every script runs identically with AI_ENABLED=false.

See Notes_AI-Design.md at the repo root for the wider design — host resolution across the Tailscale mesh, per-feature rollout tiers, and the conf-write guardrails that would have to be built before any of that is enabled.

━━━ SCHEDULING ━━━

The index tracks the repo automatically, from the only moment the corpus actually changes — a successful pull. git_pull_execute.sh re-indexes behind three gates: the pull succeeded and changed tracked files, AI_INDEX_ON_PULL=true, and AI_ENABLED=true. It is never fatal — a failed index leaves the previous one in place and the pull still reports success.

No cron entry and no DAILY_MAINTENANCE_SCRIPTS line are needed; the daily pull carries it.

An incremental run on an unchanged repo is ~70 ms, so a daily entry costs effectively nothing and a pull that changed twelve files costs a few seconds.

━━━ PROFILES ━━━

A profile is a contract plus a set of inputs. Plugin/unraid/include/ai_profiles.php is the one definition of both, read by the endpoint, the worker, the shared chat include and the Scheduler dock.

Profile	Turns	Retrieves	Notes
`varaverk`	3	yes	answers only from the index, with citations. The default.
`chat`	8	no	ordinary conversation. Holds zero capabilities, deliberately.
`code`	4	no	drafts shell for Custom Scripts; scans its own output for destructive ops
`troubleshoot`	2	yes	reasons from evidence — an open log, or one you name. May file bug reports.

Capabilities are granted per profile — retrieval, live health, run evidence, scoped log, incidents, conf lookup, bug filing, code scanning. chat holding an empty list is a guarantee, not an oversight: anything added to it stops being general chat and becomes an assistant that sometimes invents claims about this installation.

Routing out of General Chat

chat hands a question to whichever profile fits, decided by vv_ai_route_from_chat(). Ordered most specific first, because these overlap on purpose:

Question	Goes to	Why
"write me a script that prunes logs"	`code`	asked for something written
"why did the daily orch fail"	`troubleshoot`	diagnostic phrasing and something here to diagnose
"how did the daily orch go"	`varaverk`	about this install, but not a fault
"what does arr_sync.sh do"	`varaverk`	names a script, wants documentation
"why is the sky blue"	stays `chat`	diagnostic phrasing about nothing here

code is checked first because it is the only intent about a thing that does not exist yet, so nothing else can claim it — and it is anchored on the verb, which is what keeps "write me a script" apart from "what does this script do". Escalation adds capability, so a wrong escalation costs more than a missed one: anything unrecognised stays in chat, the profile that cannot invent claims about this system. The worker reverts to chat anyway if retrieval comes back empty.

The answer opens with one line naming the profile that took it, because the button still shows the one you picked and an answer arriving under a different contract otherwise reads as the assistant ignoring you.

Routing is asserted by Plugin/unraid/Tools/ai_explain_check.sh against ai_explain_fixtures.txt — every case runs through the worker's --explain mode, which stops where deterministic assembly ends and never calls the model.

This used to live in five places — history depth in the endpoint, capabilities in include/ai.php, label and depth again in JavaScript, a prompt branch in the worker, and a label map on the Scheduler page. They had already drifted: the JavaScript knew three profiles where PHP knew four. The system prompts still live in Tools/ai_chat_worker.php, because they have one reader and moving them would relocate the most delicate text in the subsystem without removing a duplicate.

━━━ CONVERSATIONS ━━━

Chats are stored server-side under AI_DATA_DIR/ai_chats/, one JSON file each, saved automatically when a turn completes and pruned to AI_CHAT_HISTORY_MAX (default 10, oldest first by creation).

There is no Save button. A conversation worth keeping is not reliably one you knew was worth keeping while you were having it.

The same store backs the AI tab and the Monitor tab's AI row, so a thread started on the dashboard is the one you carry on in the tab. Messages are re-validated per message on the way in — a stored chat is replayed into a later prompt when reopened, so an unchecked role written there would be an injection that survives a reload rather than one turn.

Reopened chats render as plain turns: sources, reasoning and timings describe one generation and are not stored, because redrawing them beside a transcript that may be continued under a different profile would be citing evidence for an answer no longer being made.

━━━ TOKEN ACCOUNTING ━━━

Every completed ask appends one row to AI_TOKEN_DB (data/ai/ai_token_history.db):

date|time|host|profile|source|prompt_tokens|completion_tokens|tok_s
2026-08-04|22:03:51|host1|varaverk|cli|2041|318|61.4

Both paths write it — this CLI (source=cli) and the WebGUI worker (source=webgui) — so the totals are not quietly the tab's alone. ai_query.sh passes --token-db and --token-host; called by hand without them, cli.js simply skips the row rather than guessing a path, because this file never reads conf itself.

Read it on the plugin's AI tab, which aggregates today / last 7 days / all time, per host. Or straight from the shell, since it is just a delimited file:

# tokens used today
awk -F'|' -v d="$(date +%F)" '$1==d {p+=$6; c+=$7} END {print p+c}' data/ai/ai_token_history.db

The host column is where the turn ran, not where the file is read. Each host writes only its own rows.

ai_token_sync.sh pulls each partner's ledger into $AI_TOKEN_CACHE_DIR/<slot>.tokens.db (/tmp/varaverk/ai/) — the same trick conf_sync.sh uses for partner confs, and it runs from INTERMEDIATE_MAINTENANCE_SCRIPTS every four hours. The tab then reads every ledger it can see, so a fleet total is a fleet total.

Pull only, no push: nothing here is needed by anyone else, and a reader that fetches its own data controls its own freshness instead of depending on the partner's cron. A partner file may only contribute rows whose host column matches its filename — a ledger copied into the wrong slot would otherwise be double-counted against a total that still looked plausible.

The cache is tmpfs with no save/restore pair, unlike the conf cache. Stale counters are worse than absent ones: absent renders as "not collected here", stale renders as fact. An unreachable partner leaves its file alone and logs at info, because a partner being down for weeks is a normal state, not an incident.

Pruning is by row count (AI_TOKEN_RETAIN_ROWS, default 20000) and happens on write, but only once the file passes a size threshold — an ordinary turn costs a stat() and an append. The CLI deliberately does not prune: duplicating a read-modify-write of the whole file in a second language is how the two drift apart.

16 KiB Raw Blame History Unescape Escape