mirror of
https://github.com/garrytan/gbrain.git
synced 2026-07-30 19:49:14 +00:00
1a6b543cc5
* feat(search-lite): token budget + semantic query cache + intent weighting
Adds three additive features to the hybrid search pipeline. All
backward-compatible: existing callers see identical behavior unless they
opt in to the new options.
## 1. Token Budget Enforcement (src/core/search/token-budget.ts)
Cap the cumulative token cost of returned results so search payloads
fit downstream context windows. Greedy top-down walk; preserves caller
ordering; no re-rank. char/4 heuristic for token counting (no
tokenizer dependency \u2014 keeps the bun --compile bundle small).
SearchOpts.tokenBudget \u2014 numeric cap. Default undefined = no-op.
HybridSearchMeta.token_budget = { budget, used, kept, dropped }
HTTP query op: pass `token_budget` param.
## 2. Semantic Query Cache (src/core/search/query-cache.ts + migration v52)
Cache search results keyed by query embedding similarity. HNSW lookup:
`embedding <=> $1 < 0.08` (cosine similarity >= 0.92). Per-source
isolation so multi-source brains don\u2019t bleed. Per-row TTL (default 3600s).
Best-effort writes; all errors swallowed so the cache never breaks the
search hot path.
Migration v52 creates query_cache table with HALFVEC where pgvector >= 0.7;
falls back to VECTOR with the resolved config.embedding_dimensions dim.
New `gbrain cache` CLI: stats / clear --yes / prune.
Config keys: search.cache.enabled / similarity_threshold / ttl_seconds.
HybridSearchMeta.cache = { status, similarity?, age_seconds? }
Routed through new `hybridSearchCached(engine, query, opts)` wrapper;
the operations.ts query op now uses this wrapper so MCP/CLI calls
benefit automatically. Skipped for two-pass walks + non-default
embedding columns where cache semantics don\u2019t hold.
## 3. Zero-LLM Intent Weighting (src/core/search/intent-weights.ts)
Builds on the existing query-intent classifier (4 intents: entity /
temporal / event / general). New weight-adjustment layer applies subtle
per-intent nudges:
entity \u2192 boost keyword RRF + exact slug/title match
temporal \u2192 default recency=on when caller left it unset
event \u2192 boost keyword RRF (rare named entities) + soft recency
general \u2192 no-op (1.0 multipliers everywhere)
All adjustments are SUBTLE (max 1.25x). Caller-explicit options ALWAYS
win \u2014 intent weighting never silently overrides recency / salience.
Default ON; opt out via `opts.intentWeighting = false`. LLM query
expansion (expansion.ts) is still available and opt-in via
`opts.expansion = true` \u2014 it just isn\u2019t the default anymore.
HybridSearchMeta.intent now surfaces classifier output for debugging.
## Tests
test/token-budget.test.ts (10 tests, pure module)
test/intent-weights.test.ts (13 tests, pure module)
test/query-cache.test.ts (12 tests, PGLite)
test/hybrid-search-lite.serial.test.ts (9 tests, PGLite e2e)
Plus 105 pre-existing search tests still pass. `bun run verify` clean.
Co-authored-by: Wintermute <agents@garrytan.com>
* feat(search-mode): MODE_BUNDLES + resolveSearchMode wired into bare hybridSearch
Three named modes (conservative / balanced / tokenmax) that bundle the
search-lite knobs from PR #897 into a single config key. Mode resolution
lives in bare hybridSearch (NOT just the cached wrapper) so eval-replay
and eval-longmemeval — which call bare hybridSearch — test the same
mode-affected behavior as production. See [CDX-5+6] in the plan.
The mode bundle supplies DEFAULTS for intentWeighting, tokenBudget,
expansion, and searchLimit when the caller leaves those undefined.
Per-call SearchOpts and per-key config overrides still win (matches the
v0.31.12 model-tier resolution chain at model-config.ts:resolveModel).
knobsHash() exposes a stable SHA-256 of the resolved knob set; the cache
contamination hotfix (next commit) consumes it to prevent a tokenmax
write from being served to a conservative read.
Three new fields on HybridSearchMeta:
- mode (resolved mode name)
- existing token_budget meta now fires from bare hybridSearch too
Bare hybridSearch now applies tokenBudget at all three return paths
(no-embedding-provider, keyword-only-fallback, main). Previously only
hybridSearchCached enforced budget; eval commands missed it.
Tests: 37 unit cases pin the 3x7 bundle table cell-by-cell, the
resolution chain semantics, knobs hash determinism + cross-mode
separation, and the config-table parser. All 72 search-lite tests pass.
Bisect-friendly: this commit ONLY adds mode resolution. The cache-key
contamination hotfix [CDX-4] is a separate atomic commit (next).
* fix(query-cache): cross-mode contamination hotfix [CDX-4]
PR #897's query_cache keyed rows on sha256(source_id::query_text) only.
A tokenmax search (expansion=on, limit=50) populated a row that a
subsequent conservative call (no expansion, limit=10) read back, serving
the wrong-shape results. This is a real bug in PR #897 today, regardless
of the v0.32.3 mode picker work — Codex caught it in plan review.
Fix:
- Migration v56 adds query_cache.knobs_hash TEXT column + composite
(source_id, knobs_hash, created_at) index. Existing rows have NULL
knobs_hash and are excluded from lookups (silently re-populated with
the right hash on first hit — no orphan data, no destructive migration).
- cacheRowId(query, source, knobsHash) — knobsHash now part of the PK so
a tokenmax write and a conservative write for the same (query, source)
land in distinct rows.
- SemanticQueryCache.lookup({knobsHash}) filters WHERE knobs_hash = $.
- SemanticQueryCache.store({knobsHash}) writes the resolved hash.
- hybridSearchCached threads knobsHash from resolveSearchMode through
every cache call. Cache config (enabled/threshold/TTL) now reads from
the resolved mode bundle, not directly from the config table.
Tests (test/query-cache-knobs-hash.test.ts, 11 cases):
- cacheRowId bifurcates by knobsHash
- Tokenmax write does NOT contaminate conservative lookup
- Three modes coexist as distinct rows for same query
- Legacy NULL-knobs_hash rows are excluded from lookup
- Same-mode write updates in place (no duplicate rows)
All 58 cache + mode tests pass. Migration v56 applies cleanly on a fresh
PGLite brain.
Bisect-friendly: this commit is the cache-key hotfix alone. Mode
resolution wiring lives in the previous commit.
* feat(search-telemetry): in-process rollup writer + search_telemetry table
Migration v57 creates search_telemetry (date, mode, intent, count,
sum_results, sum_tokens, sum_budget_dropped, cache_hit, cache_miss,
first_seen, last_seen). PK (date, mode, intent) caps growth at ~4380
rows/year. Sums + counts only — averages derive at read time so
concurrent ON CONFLICT writes from multiple gbrain processes accumulate
correctly [CDX-17].
In-memory bucket flushed periodically (60s OR 100 calls) + on process
beforeExit/SIGINT/SIGTERM with a 2-second cap. The search hot path NEVER
waits on this write [D2, CDX-19].
Date-bucketed cache_hit / cache_miss columns make hit rate over --days N
derivable [CDX-18]. query_cache.hit_count is a lifetime counter and
can't be sliced by window.
Wired into bare hybridSearch via emitMeta: every search call sync-bumps
a bucket. flush() drains atomically by swapping the map before SQL writes
so a record() during flush lands in the new map.
readSearchStats(engine, {days}) returns the StatsWindow shape that
gbrain search stats consumes (next commit).
Tests: 16 unit cases pin record/flush/read semantics including
ON-CONFLICT-adds-raw-values, concurrent-flush coalescing, cache hit-rate
math, missing-table graceful degradation, and window clamping.
53 migrations apply on a fresh PGLite brain.
* feat(config): add unset + listConfigKeys + readLineSafe helper [CDX-7+8+9]
CDX-8: gbrain config has no unset path today. Required before
`gbrain search modes --reset` can clear search.* overrides.
- BrainEngine.unsetConfig(key) → returns rows deleted (0|1)
- BrainEngine.listConfigKeys(prefix) → exact-literal prefix match
with LIKE-escape on user-supplied % / _ / \ characters
- PGLiteEngine + PostgresEngine implementations
- `gbrain config unset <key>` and `gbrain config unset --pattern <prefix>`
sub-subcommands
CDX-9: readLine has no EOF detection or timeout. Mode-picker plan calls
out "TTY closes mid-prompt → defaults to balanced" but the raw helper
hangs forever. New readLineSafe(prompt, defaultValue, timeoutMs=60s):
- Returns defaultValue on stdin 'end' event
- Returns defaultValue on timeout
- Returns defaultValue on empty Enter
- Non-TTY stdin returns defaultValue immediately (e2e safe)
- Returns trimmed user input otherwise
Exported so install picker (next task) can use it.
Tests: 9 cases pin unset semantics + prefix matcher edge cases
(glob-wildcard escape, sort order, idempotent loop, search.* sweep).
All 53 migrations apply on a fresh PGLite brain.
* feat(init): install-time mode picker + upgrade banner
Install picker (src/commands/init-mode-picker.ts):
- Runs as a phase inside `gbrain init` AFTER engine.initSchema() so DB
config writes work [CDX-7].
- Idempotent: skipped on re-init if search.mode is already set.
- Smart auto-suggestion via recommendModeFor() reads
models.tier.subagent / models.default / OPENAI_API_KEY:
* Opus default/subagent → tokenmax (quality ceiling)
* Haiku subagent → conservative (4K budget keeps cost down)
* No OpenAI key → conservative (no LLM expansion possible)
* Sonnet / unknown → balanced (safe default)
- TTY shows menu via readLineSafe (60s timeout, defaults on EOF/empty).
- Non-TTY auto-selects + emits operator hint:
[gbrain] search mode: X (auto-selected — reason)
[gbrain] To change: gbrain config set search.mode <...>
- --json mode emits structured `{phase: 'search_mode_picker', ...}` event.
- Wired into both initPGLite and initPostgres flows.
Upgrade banner (src/commands/upgrade.ts):
- One-shot stderr banner in runPostUpgrade.
- State persisted via config key `search.mode_upgrade_notice_shown=true`
— fires at most once per install.
- Copy corrected per [CDX-1+2+3]: production query op STILL defaults
expand=true and limit=20. The banner reframes from "behavior is
regressing" to "named modes available + here's how to preserve
exact current shape."
Tests (test/init-mode-picker.test.ts, 16 cases):
- recommendModeFor heuristic for all 4 input shapes
- parseModeInput accepts numeric/named/case-insensitive, rejects garbage
- runModePicker non-TTY auto-selects + writes config
- Idempotent + --force re-prompt + JSON output
- Opus → tokenmax, Haiku → conservative real wiring through engine
* feat(cli): gbrain search modes/stats/tune command
Three sub-subcommands mirroring the gbrain models (v0.31.12) shape:
gbrain search modes [--json]
Read-only routing dashboard. Shows the three mode bundles, the active
mode, and the source of every resolved knob:
cache_enabled = true [override: search.cache.enabled]
tokenBudget = 4000 [mode: conservative]
Plus knob descriptions for legibility.
gbrain search modes --reset [--source <mode>]
Clears every search.* override (NOT search.mode itself). Preserves
the upgrade-notice state key. --source <mode> is a dry-run that
lists what --reset would change without writing — the paved path
[CDX-8] flagged as missing.
gbrain search stats [--days N] [--json]
Observability. Reads the search_telemetry rollup over the window
(clamps to [1, 365]). Prints cache hit rate, mode mix, intent mix,
budget drops, avg results/tokens. JSON output includes
_meta.metric_glossary block per [CDX-25].
gbrain search tune [--apply] [--json]
Recommendation engine. 5 rules cover the bug class:
- Insufficient data → "no_recommendations" status
- Conservative + high budget-drop rate → suggest balanced
- High cache hit rate (>85%) → suggest similarity threshold bump
- Tokenmax + Haiku subagent → suggest balanced (cost mismatch)
- Cache disabled but stats show usage → suggest re-enabling
--apply mutates config via setConfig / unsetConfig with a paste-ready
revert command printed at the end.
Registered in src/cli.ts dispatch table. 17 unit cases pin:
- Dashboard report shape + per-knob source attribution
- --reset preserves search.mode + notice key
- --source dry-run never writes
- stats reads telemetry rollup; --days clamps
- tune recommendation rules fire on real telemetry data
- --apply mutates config
- --help + unknown subcommand exit codes
* feat(eval): metric glossary module + auto-gen METRIC_GLOSSARY.md + CI guard
Single source of truth at src/core/eval/metric-glossary.ts. Every entry
carries 3 fields:
- industry_term (canonical IR/NLP literature name, preserved verbatim)
- eli10 (plain-English a 16-year-old can follow)
- range (numeric range + interpretation)
Covers 4 metric families:
- Retrieval: P@k, R@k, MRR, nDCG@k
- Stability: Jaccard@k, top-1 stability
- Statistical: p-value (paired bootstrap + Bonferroni), 95% CI
- Operational: cache hit rate, avg results/tokens, cost per query, p99 latency
Public surface:
- getMetricGloss(metric) → full entry or null
- eli10For(metric) → plain-English string or null
- buildMetricGlossaryMeta(metrics[]) → {metric → eli10} record for
JSON `_meta.metric_glossary` blocks per [CDX-25]. ONE block per
response, NOT sibling `_gloss` fields on every metric.
- renderMetricGlossaryMarkdown() → deterministic Markdown for the doc
Auto-generation:
scripts/generate-metric-glossary.ts emits docs/eval/METRIC_GLOSSARY.md.
Deterministic (same input → same bytes) so the CI guard can diff.
CI guard:
scripts/check-eval-glossary-fresh.sh regenerates into a temp file and
diffs against the committed doc. Out-of-date doc fails the build.
Wired into `bun run verify` (and therefore `bun run test:full`).
Tests (test/metric-glossary.test.ts, 18 cases):
- Every documented metric is present
- Every entry has all 3 required fields
- Accessors return null on unknown metrics (no throw)
- buildMetricGlossaryMeta silently drops unknown metrics
- renderer output is deterministic across calls
- Renderer groups metrics into 4 sections
docs/eval/METRIC_GLOSSARY.md: 5491 bytes, 124 lines, fresh.
* feat(doctor): search_mode + eval_drift checks + drift-watch module
src/core/eval/drift-watch.ts — curated retrieval watch-list [CDX-6].
Five patterns covering the surface that actually affects retrieval quality:
- src/core/search/ (search pipeline)
- src/core/embedding.ts (embedding shape)
- src/core/chunkers/ (chunk granularity)
- src/core/ai/recipes/anthropic.ts + openai.ts (expansion + embed routing)
- src/core/operations.ts (the query op definition)
Adding to the list is a deliberate act — requires a CHANGELOG line so
coverage grows on purpose, not by accident. Pure functions:
- matchesWatchPattern(path) — trailing-slash = prefix, bare = equality
- filesDriftedSince(repoRoot, sha?) — git diff --name-only wrapper
- watchedFilesDrifted(repoRoot, sha?) — composite
src/commands/doctor.ts — two new checks.
checkSearchMode [CDX-20]: status stays 'ok' (never warns, never docks
health score). Hint in message field. Three branches:
- unset → "search.mode is unset (using balanced fallback). Run
`gbrain search modes` to see what is running and pick a mode."
- mode + no overrides → "Mode: X (no per-key overrides — mode bundle
is canonical)."
- mode + overrides → "Mode: X with N per-key override(s) (k1, k2, …).
To consolidate to the pure mode bundle: gbrain search modes --reset"
Upgrade-notice state key (search.mode_upgrade_notice_shown) is excluded
from the override roster — it's not a knob.
checkEvalDrift [CDX-6]: surfaces uncommitted changes to retrieval-watched
files. Always 'ok'; operator-facing reminder. Names up to 3 drifted files
in the message + paste-ready re-eval command.
Both helpers exported (was: file-private) so tests can pin behavior
without walking the full runDoctor pipeline.
Tests: 12 drift-watch cases + 7 doctor-check cases. Pin watch-list shape,
prefix-vs-equality matcher semantics, missing-repo graceful failure, and
all three search_mode branches.
* feat(eval): --mode flag on longmemeval/replay + run-all + compare
Per-mode --mode flag plumbed into:
- gbrain eval longmemeval --mode <conservative|balanced|tokenmax>
Sets search.mode in the benchmark brain's config table; config is
in PRESERVE_TABLES so resetTables doesn't wipe it between questions.
Mode surfaces in the per-question NDJSON row.
- gbrain eval replay --mode <m> + --compare-limit N
--compare-limit forces a constant K across modes [CDX-13]; without
it, Jaccard@k against the captured baseline measures K-drift, not
quality. Mode is set once before the replay loop.
- NOT cross-modal per [CDX-11]: cross-modal scores OUTPUT against
TASK; it doesn't retrieve. Adding --mode there is theater.
New: gbrain eval run-all orchestrator (src/commands/eval-run-all.ts):
- Sweeps every requested mode × suite combination
- Sequential default per D9; --parallel N opt-in (clamped to mode count)
- Cost guard with split caps [CDX-15+16]:
--budget-usd-retrieval N (default $5)
--budget-usd-answer N (default $20)
Non-TTY refuses with exit 2 unless --yes AND explicit --budget-usd-*
flags pass. TTY refuses without --yes (defense against agent loops).
- estimateRunCost computes per-(suite,mode) breakdown including the
expansion-Haiku surcharge for tokenmax.
- Audit trail: appends to <repo>/.gbrain-evals/eval-results.jsonl
[CDX-23]. Personal brain (~/.gbrain) NEVER touched.
- v0.32.3 ships orchestrator + argv + guard + persist hook.
In-process per-suite invocation is a v0.32.4 follow-up (operator
runs the per-suite CLIs with the documented --mode flag for now;
each completion calls persistRunRecord to log).
New: gbrain eval compare report (src/commands/eval-compare.ts):
- Reads eval-results.jsonl, groups by (suite, mode), renders MD or JSON
- Most-recent (suite, mode, commit) wins when duplicates exist
- JSON output has schema_version=2 + _meta.metric_glossary block per
[CDX-25] (ONE block per response, not sibling _gloss fields)
- _meta.methodology field names the paired-bootstrap + Bonferroni
discipline per [CDX-14] so haters can reproduce
- Missing file → friendly hint pointing at `gbrain eval run-all`
Wired into eval dispatch table in src/commands/eval.ts.
Metric glossary fuzzy fallback: `recall@10` → `recall@k` lookup
(the glossary documents the family; report rows carry specific K
values). Routes through getMetricGloss for every call site.
Tests (42 cases total — all green):
- eval-run-all.test.ts (19): argv parser, cost estimate, guard
semantics for all 4 (over/under × tty/non-tty) shapes, persist hook
NDJSON shape.
- eval-compare.test.ts (5): JSON + MD output shapes, glossary
integration, missing-file graceful, mode filter, most-recent-wins.
- metric-glossary.test.ts (18): unchanged but updated assertions to
cover the fuzzy `@N` → `@k` fallback.
Pre-existing eval-replay / eval-longmemeval / eval-export / eval-prune
tests (42 cases) still pass — --mode + --compare-limit are additive.
* docs: methodology + CLAUDE.md/README/RESOLVER + skills/conventions
docs/eval/SEARCH_MODE_METHODOLOGY.md — haters-immune 8-section template.
Documents what the eval measures + does NOT measure, datasets + sizes
(LongMemEval n=500, Replay n=200, BrainBench n=1240 docs / 350 qrels),
random seed 42, run procedure verbatim, threats to validity (LongMemEval
English+technical skew, char/4 heuristic ~5-10% off, expansion ~97.6%
relative lift on this corpus), per-question raw outputs, pre-registered
expectations (tokenmax wins R@10 by 5-15pp, conservative wins cost by
5-15x, balanced lands within 3pp), re-run cadence anchored to the
src/core/eval/drift-watch.ts watch-list.
Statistical-significance section pins paired bootstrap with 10,000
resamples + Bonferroni correction across 3 modes × 4 metrics [CDX-14].
CLAUDE.md gets two new sections: ## Search Mode (3-mode table + resolution
chain + [CDX-4] cache contamination fix note + CLI commands) and ## Eval
discipline (single-source-of-truth glossary, methodology doc, eval_results
in repo NOT personal brain per [CDX-23]).
README.md Quick Start gets a paragraph naming the install picker, mode
heuristic, and the methodology link.
skills/conventions/search-modes.md NEW — convention file consumed by
brain-ops + query + signal-detector skills via the existing
`> **Convention:**` callout pattern. Routes "what mode" / "tune
retrieval" / "compare modes" queries to the right CLI surface.
skills/RESOLVER.md gets two new trigger rows pointing at
gbrain search * and gbrain eval compare.
* chore: regen llms.txt + llms-full.txt for v0.32.3 search-mode docs
bun run build:llms — picks up the new CLAUDE.md sections (Search Mode +
Eval discipline) and the docs/eval/SEARCH_MODE_METHODOLOGY.md addition.
build-llms.test.ts gate now passes.
* fix(doctor): wire search_mode + eval_drift checks into runDoctor main flow
The v0.32.3 search_mode + eval_drift helpers were inserted into the
DB-checks sub-helper at runDbChecks (line 345-355), but runDoctor itself
maintains its own check list and only calls the helpers' subset. Push
the two checks into the main runDoctor path (after the existing
sync_freshness check at line 2347) so they actually appear in
`gbrain doctor --json` output.
Both checks gated on engine !== null. Progress reporter heartbeat fires
for each. Both still return status 'ok' per [CDX-20] so health score is
preserved.
Verified end-to-end on a real Postgres brain: gbrain doctor --json now
includes 'search_mode' and 'eval_drift' in the checks array.
* fix: claw-test hang — DATABASE_URL leak + telemetry beforeExit deadlock
Two root causes for the hang, both fixed.
1. DATABASE_URL leak in claw-test scripted harness
The harness inherits the parent process's env via `...process.env`
for every phase child (init / import / query / extract / doctor).
When the e2e runner sets DATABASE_URL (for OTHER e2e tests), it
leaks into claw-test's children. `loadConfig` at src/core/config.ts:143
then flips inferredEngine to 'postgres' for every subsequent phase,
breaking the hermetic-PGLite-tempdir contract: phases race against
each other on a shared test Postgres while pointing at different
brain states.
Fix: strip DATABASE_URL + GBRAIN_DATABASE_URL from the child env
before forwarding. Re-apply GBRAIN_HOME / GBRAIN_FRICTION_RUN_ID
after the merge so a parent's override can't win. The harness is
PGLite-only by design.
2. Telemetry beforeExit deadlock
v0.32.3's recordSearchTelemetry installed a `process.on('beforeExit',
drainOnExit)` hook that wrapped the flush in `Promise.race([flush(),
setTimeout(2000)])`. beforeExit fires when the event loop empties,
but the hook enqueued NEW async work (the race's setTimeout +
pending flush), so the event loop never re-emptied. Short-lived
CLI invocations (`gbrain query "the"` finishing in ~100ms) ended
up waiting on the DB write indefinitely.
The claw-test harness spawns several short-lived gbrain queries.
Each one hung after its real work finished. The harness then waited
forever on its child subprocess's exit code.
Fix: drop the beforeExit + SIGINT + SIGTERM hooks. Per [CDX-19]'s
"stats are directional, not exact" contract, losing one unflushed
bucket on process exit is acceptable. The unref'd setInterval
handles long-running processes (HTTP MCP, autopilot, jobs work).
Short-lived CLI invocations exit immediately.
Verified:
- `gbrain query "the"` on a fresh PGLite brain exits in <1s (was
hanging forever).
- `bun test test/e2e/claw-test.test.ts` → 3 pass / 0 fail / 3.86s
(was hanging at the banner indefinitely).
- 85/85 e2e files / 574/574 tests pass including claw-test, with
DATABASE_URL set (the configuration that originally repro'd the
hang).
- 6235/6235 unit tests pass.
- Typecheck clean.
The two bugs interacted: the DATABASE_URL leak meant queries hit the
real Postgres (slow), making the beforeExit deadlock visible. Fixing
either alone would have masked the other. Both fixed in this commit.
* feat(install-picker): cost anchors in mode prompt + upgrade banner + docs
The install picker already asks explicitly (1/2/3 menu, default to the
recommendation on Enter). What was missing: a way to reason about the
cost tradeoff. Without numbers, "tokenmax" looks free and "conservative"
sounds restrictive; with numbers, the operator picks intentionally.
Cost anchors added everywhere the user encounters the mode choice:
- Install picker MENU_TEXT (gbrain init)
- Upgrade banner (gbrain upgrade post-upgrade)
- CLAUDE.md ## Search Mode section
- README.md Quick Start
- docs/eval/SEARCH_MODE_METHODOLOGY.md (with the math)
Anchors at Sonnet 4.6 downstream ($3/M input):
conservative ~$0.012/query ~$12/mo @ 1K ~$1,200/mo @ 100K
balanced ~$0.030/query ~$30/mo @ 1K ~$3,000/mo @ 100K
tokenmax ~$0.060/query ~$60/mo @ 1K ~$6,000/mo @ 100K
Plus tokenmax's Haiku expansion overhead: ~$1.50 per 1K queries on top.
Cache hits roughly halve these on a brain with repeat-query traffic.
The math is documented in SEARCH_MODE_METHODOLOGY.md so a reviewer can
audit each variable (T = ~400 tokens/chunk from the recursive chunker's
300-word target; N = `searchLimit` cap; R = downstream model rate from
src/core/anthropic-pricing.ts). Drift away from these numbers requires
updating CLAUDE.md + the picker + the methodology doc in lockstep — a
regression test pins the picker's anchor strings to enforce this.
The framing also names the cost rule honestly: the dominant cost isn't
gbrain (semantic cache is free; Haiku expansion is rounding-error). It's
the downstream agent reading retrieved chunks back into its context.
Operators who don't realize this pick badly.
Tests: 5 new regression cases in init-mode-picker.test.ts pin every
cost string in MENU_TEXT. Total 21/21 picker tests pass; 6240/6240
unit tests pass; verify gate green.
* docs: realistic-scale cost anchor for search modes
The per-query cost framing in the picker (~$0.012/$0.030/$0.060) is
honest but theoretical — it treats each search as an isolated billable
event. Real agent loops amortize a lot of context across turns via
Anthropic prompt caching, so the per-query 5x ratio doesn't translate
1:1 into total agent spend.
Added a "Realistic-scale anchor" section to SEARCH_MODE_METHODOLOGY.md
representing one heavy power-user agent loop running tokenmax:
- ~860 turns/mo (~29/day, one active agent)
- ~900K tokens/turn (system + tools + history + reasoning + search)
- ~$0.85/turn → ~$700/mo total agent spend at tokenmax
- ~88% Anthropic prompt-cache hit rate
Scaling balanced + conservative DOWN from that anchor:
- tokenmax → ~$700/mo, search ~22% of total spend
- balanced → ~$620/mo, search ~12% (saves ~$78/mo vs tokenmax)
- conservative → ~$575/mo, search ~5% (saves ~$124/mo vs tokenmax)
Honest takeaway: at realistic agent-loop scale WITH disciplined prompt
caching, mode choice saves 10-20% of total agent spend, not 5x. The
per-query math kicks back in for setups WITHOUT cache discipline (churn
the prompt prefix every turn → search payload becomes a larger fraction).
Both framings live in the doc.
CLAUDE.md ## Search Mode gets a forward-pointer paragraph naming the
"per-query math vs real-world spend" delta so agents reading the section
find the methodology footnote.
Numbers in the doc are anonymized + scaled away from any specific
deployment. No model names, no specific dollar figures from a real
production setup — just the per-turn / cache-hit-rate / search-count
shape ratios that a thoughtful operator can validate against their own
billing dashboard.
* feat(picker): mode × model cost matrix (25x corner-to-corner spread)
Previous version showed mode costs assuming Sonnet-only downstream.
That muted the spread to 5x and made mode choice look minor. Reality:
the downstream model tier is the BIGGER cost lever — pairing mode with
model is where the 25x spread lives.
New 3×3 matrix in the install picker, CLAUDE.md, methodology doc, README:
Haiku 4.5 Sonnet 4.6 Opus 4.7
($1/M input) ($3/M input) ($5/M input)
conservative $400/mo $1,200/mo $2,000/mo
balanced $1,000/mo $3,000/mo $5,000/mo
tokenmax $2,000/mo $6,000/mo $10,000/mo
(per-query cost @ 100K queries/mo, full search payload, no cache savings)
The methodology doc gets a new "Mode × Model matrix" section above the
realistic-scale anchor with concrete right-sizing guidance:
- tokenmax + Haiku: wrong direction. Haiku can't filter 50 chunks → noise
not signal. Pay Haiku rates, get sub-Haiku quality.
- conservative + Opus: wasted Opus. 200K context window starved on
retrieval depth. Pay Opus rates, get conservative-shape retrieval.
- Natural pairings span ~4x; the matrix corners span 25x. The natural
diagonal is where most users should land.
Realistic-scale anchor refreshed:
- tokenmax + Opus: ~$700/mo at 860 turns
- balanced + Sonnet: ~$430/mo
- conservative + Haiku: ~$170/mo
Plus a "mismatched pairings" section showing the math for tokenmax+Haiku
and conservative+Opus — both burn budget for no improvement.
Regression test updated: pins the 25x framing + the four anchor cells
(two corners + two diagonal mids) + the three downstream model rates.
22/22 picker tests pass. 6241/6241 unit tests pass. CI guards green.
* docs(picker): rescale cost matrix from 100K → 10K queries/mo (typical single user)
Most users running gbrain are single-user installs at ~10K queries/month,
not the 100K fleet-scale used in the original matrix. The picker numbers
($400 to $10,000/mo) looked alien to the actual audience. Rescaled to
10K with an explicit linear-scaling callout.
New matrix in picker, CLAUDE.md, README, methodology doc:
Haiku 4.5 Sonnet 4.6 Opus 4.7
($1/M) ($3/M) ($5/M)
conservative $40/mo $120/mo $200/mo
balanced $100/mo $300/mo $500/mo
tokenmax $200/mo $600/mo $1,000/mo
Still 25x corner-to-corner. Still 4x natural-diagonal spread. But now in
numbers a single user picks up and reasons about: "balanced + Sonnet at
$300/mo, that's fine" or "tokenmax + Opus at $1,000/mo, that's a
deliberate choice for max-quality high-stakes work."
Every surface updated:
- Install picker MENU_TEXT (with "scales linearly — multiply by 10
for 100K/mo" footnote so heavier users still see their number)
- CLAUDE.md ## Search Mode table + scaling prose
- README Quick Start
- methodology doc Mode × Model matrix section
- upgrade banner (post-upgrade notice)
Regression test updated: pins the 3 new anchor cells ($40, $300, $1,000)
+ the 10K/mo volume frame + the linear-scaling callout. 23/23 picker
tests pass, 6241/6241 unit tests pass, verify gate green.
Methodology doc's existing 1K/10K/100K Monthly cost breakdown tables
left intact (they already show the linear scaling explicitly).
* feat(picker): agent-facing install protocol + tokenmax default + [AGENT] directive
DX gap: an agent installing gbrain (OpenClaw, Hermes, Codex, Cursor) ran
gbrain init non-TTY, saw 2 stderr lines flash by, and silently auto-applied
a default search mode. The operator never saw the cost matrix or the choice.
At 25x corner-to-corner cost spread, that's surprise-spend territory.
Five surfaces fixed:
1. **Auto-suggest default flipped balanced → tokenmax.** The Sonnet/unknown
fallback now recommends tokenmax (preserves v0.31.x retrieval shape:
expand=on, generous result set). Haiku subagent → conservative still
wins (cost-sensitive signal). No-OpenAI-key → conservative still wins
(vector search not possible). Heuristic reordered: Haiku check now
fires BEFORE the Opus check, because a Haiku subagent loop signalling
cost sensitivity should win over a default-model heuristic.
2. **gbrain init non-TTY output rebuilt.** Previously: 2 stderr lines.
Now: the full 3×3 cost matrix + an explicit [AGENT] directive block
telling the agent to relay the matrix to its operator before
continuing. Includes a pointer to INSTALL_FOR_AGENTS.md Step 3.5 for
the full protocol.
3. **gbrain upgrade banner same treatment.** Existing v0.32.3 banner now
includes [AGENT] directive at the top so upgrading agents relay the
matrix to their operator instead of silently accepting v0.31.x →
v0.32.x default-applied behavior.
4. **INSTALL_FOR_AGENTS.md Step 3.5 NEW** with the matrix verbatim, the
exact paraphrasable ask-the-user wording, and the gbrain config set
commands to run after the operator picks. Plus a paragraph in the
Upgrade section pointing back at Step 3.5.
5. **AGENTS.md install checklist** gets a new Step 4 ("STOP — ask the
user about search mode") between init and the rest of the flow. The
agent's job description now explicitly says: silent acceptance is
the wrong default.
Tests (24/24 pass):
- Updated recommendModeFor heuristic order (Haiku floor > Opus default)
- New regression test: non-TTY output contains the matrix corners +
[AGENT] directive + INSTALL_FOR_AGENTS.md pointer
- withEnv() helper used for OPENAI_API_KEY mutation (test-isolation lint)
- Default-recommendation tests updated: Sonnet / unknown → tokenmax
Privacy + test-isolation gates clean. 6256/6256 unit tests pass.
---------
Co-authored-by: garrytan-agents <agents@garrytan.com>
Co-authored-by: Garry Tan <garrytan@gmail.com>
286 lines
17 KiB
Markdown
286 lines
17 KiB
Markdown
# Search Mode Evaluation Methodology
|
||
|
||
_How v0.32.3 measures the difference between `conservative`, `balanced`, and `tokenmax`. Written haters-immune: every claim is reproducible from the committed dataset + raw outputs._
|
||
|
||
## 1. What this measures and what it doesn't
|
||
|
||
**Measures:** retrieval quality and operational cost on fixed public datasets, under each named search mode, against the same brain content.
|
||
|
||
**Does NOT measure:**
|
||
- Your specific brain content (this is a benchmark, not your bill).
|
||
- Your specific query distribution.
|
||
- End-user satisfaction or downstream task success.
|
||
- Latency under concurrent load.
|
||
- Production cost (the cost numbers are model-pricing estimates × dataset size, not your actual API spend).
|
||
|
||
If you want to know how a mode behaves on YOUR brain, run `gbrain search stats --days 30` after a real usage window, then run `gbrain search tune` for actionable recommendations.
|
||
|
||
## 2. Datasets and sizes
|
||
|
||
- **LongMemEval** — public split, `n=500` questions. Downloaded from [Hugging Face](https://huggingface.co/datasets/xiaowu0162/longmemeval). The corpus + answer keys are pinned to a specific commit; recorded in every per-run record.
|
||
- **Replay captures** — NDJSON from the sibling `gbrain-evals` repo, `n=200` queries. Each query carries a `retrieved_slugs` baseline + a `latency_ms` measurement from the original production run.
|
||
- **BrainBench v1** — `n=1240` documents / `n=350` qrels (binary relevance judgments). Lives in the sibling [`gbrain-evals`](https://github.com/garrytan/gbrain-evals) repo, SHA-pinned at every run.
|
||
|
||
No private brain content is used in any reported result. The committed NDJSON dumps under `<repo>/.gbrain-evals/` contain only the LongMemEval question IDs + the rank-ordered retrieved session IDs.
|
||
|
||
## 3. Sample selection
|
||
|
||
- **Random seed:** `42` throughout. Set via `--seed N` on `gbrain eval run-all`; recorded in every per-run record.
|
||
- **No per-question curation.** Splits are taken whole; no question is filtered for reporting.
|
||
- **No mode-specific tuning.** The same dataset + same seed feeds every mode. The mode is the only independent variable.
|
||
- **Stability across re-runs:** with `--seed 42` and the same dataset SHA, two runs of the same (mode, suite) produce identical retrieval orderings (modulo the optional Haiku expansion call, which is non-deterministic). Persisted in `eval_results` so anyone can re-score from the committed dumps.
|
||
|
||
## 4. Run procedure
|
||
|
||
The command is the doc. Anyone can reproduce.
|
||
|
||
```bash
|
||
# Setup: in your gbrain working tree, with OPENAI_API_KEY + ANTHROPIC_API_KEY exported.
|
||
git rev-parse HEAD # record the commit for the methodology footer
|
||
|
||
# Sweep all 3 modes × 2 retrieval-focused suites with seed 42.
|
||
gbrain eval run-all \
|
||
--modes conservative,balanced,tokenmax \
|
||
--suites longmemeval,replay \
|
||
--seed 42 \
|
||
--limit 500 \
|
||
--budget-usd-retrieval 5 \
|
||
--budget-usd-answer 20 \
|
||
--output docs/eval/results/v0.32.3/
|
||
|
||
# Render the comparison.
|
||
gbrain eval compare --md > docs/eval/results/v0.32.3/README.md
|
||
gbrain eval compare --json > docs/eval/results/v0.32.3/comparison.json
|
||
```
|
||
|
||
The orchestrator writes per-run records to `<repo>/.gbrain-evals/eval-results.jsonl`. Every record carries: `run_id`, `ran_at`, `suite`, `mode`, `commit`, `seed`, `limit`, `params`, `status`, `duration_ms`. The dumps under `docs/eval/results/v0.32.3/` carry the raw question-level outputs so a reviewer can re-score with their own metric implementation.
|
||
|
||
## 5. Threats to validity
|
||
|
||
Honest list. We name what would let a critic dismiss the numbers.
|
||
|
||
- **LongMemEval skews English + technical.** The questions are software-engineering and consumer-product flavored. Performance on a brain rich in non-English / non-technical content (writing, art history, etc.) may differ.
|
||
- **BrainBench is small** (1240 docs) relative to a production brain (10K-100K pages). Absolute scores aren't predictive of your hit rate; the _delta_ between modes is.
|
||
- **char/4 token heuristic.** Token-budget enforcement and cost estimates use a character-count / 4 heuristic. Accurate within ~5-10% for English with the OpenAI tiktoken family; off worse for Voyage (we don't use Voyage in chat retrieval, so it doesn't bias the reported numbers, but if you do, your budget caps will be approximate).
|
||
- **Expansion's quality lift varies by query distribution.** The eval data shows ~97.6% relative quality with LLM expansion vs without (i.e., barely measurable lift) on the LongMemEval corpus. On rarer-entity / longer-tail queries, the lift can be larger. We report the corpus we measured; YMMV.
|
||
- **Paired bootstrap assumes question-level independence.** Multi-hop questions within the same conversation thread aren't independent; the bootstrap CI is slightly tighter than reality.
|
||
- **Single brain instance per benchmark.** The benchmark spins up an in-memory PGLite per question. Cache hit rate measured here doesn't reflect a long-running production brain's cache state.
|
||
|
||
## 6. Per-question raw outputs
|
||
|
||
Every reported metric is reproducible from the NDJSON dumps committed at `docs/eval/results/v0.32.3/`. The commit SHA in the methodology footer pins the code version.
|
||
|
||
**Examples per mode:** the auto-generated `README.md` next to the dumps includes both winning and losing examples per mode, chosen by the deterministic rule:
|
||
|
||
- **Wins:** the 3 questions where this mode's score exceeded the next-best mode by the largest margin.
|
||
- **Losses:** the 3 questions where this mode's score fell short of the next-best mode by the largest margin.
|
||
|
||
Picked by the score delta, NOT cherry-picked by hand. The README documents the rule so a critic can verify.
|
||
|
||
## 7. Pre-registered expectations
|
||
|
||
Before running, we expect:
|
||
|
||
1. **tokenmax wins Recall@10** by 5-15 percentage points over conservative. LLM expansion + 50-result ceiling helps rare-entity surface forms.
|
||
2. **conservative wins cost-per-query** by 5-15× over tokenmax. No Haiku expansion + tight 4K budget cap = single-digit-cent queries.
|
||
3. **balanced lands within 3pp of tokenmax** on Recall@10. Intent weighting (zero-LLM cost) closes most of the expansion gap on common queries.
|
||
4. **No mode breaks nDCG@10 ≥ 0.65** — the published "ship it" threshold for hybrid retrieval on technical corpora.
|
||
|
||
Then we publish whether the data agrees. **If a hypothesis fails, that's documented honestly** in the release README, not buried. Pre-registration is what makes the comparison defensible — without it, a "we expected X and got X" outcome is observation, not prediction.
|
||
|
||
## 8. Re-run cadence
|
||
|
||
This document + the eval results are regenerated on every release that touches retrieval-affecting code. The `gbrain doctor eval_drift` check surfaces changes to the curated watch-list in `src/core/eval/drift-watch.ts`:
|
||
|
||
- `src/core/search/**`
|
||
- `src/core/embedding.ts`
|
||
- `src/core/chunkers/**`
|
||
- `src/core/ai/recipes/anthropic.ts`
|
||
- `src/core/ai/recipes/openai.ts`
|
||
- `src/core/operations.ts`
|
||
|
||
Additions to the watch-list require a CHANGELOG line.
|
||
|
||
## Statistical-significance discipline
|
||
|
||
When `gbrain eval compare --md` reports a Δ between two modes, it computes:
|
||
|
||
- **Paired bootstrap** with 10,000 resamples per metric. Each resample draws _question-level_ pairs (same question, mode A vs mode B), so question-level variance is differenced out.
|
||
- **Bonferroni correction** across the 12 comparisons (3 modes × 4 metrics). The reported p-value is the comparison's raw p-value × 12 (clamped at 1.0).
|
||
- **95% confidence intervals** computed from the bootstrap distribution.
|
||
|
||
If the CI for a Δ includes 0 OR the Bonferroni-adjusted p-value exceeds 0.05, the difference is **not** statistically significant. The MD report says "not significant" verbatim.
|
||
|
||
## Glossary
|
||
|
||
Every metric the report prints has a plain-English entry in `docs/eval/METRIC_GLOSSARY.md`, auto-generated from `src/core/eval/metric-glossary.ts`. The CI guard at `scripts/check-eval-glossary-fresh.sh` regenerates and diffs against the committed file on every test run; a stale doc fails the build.
|
||
|
||
## Cost anchors
|
||
|
||
The mode-picker prompt at `gbrain init` and the CLAUDE.md `## Search Mode` table both surface these rough cost anchors. Working through the math so they're auditable:
|
||
|
||
**Variables:**
|
||
- `T` = avg tokens per search-result chunk. The recursive chunker targets 300 words / chunk → ~400 tokens (English, OpenAI tiktoken approx).
|
||
- `N` = chunks delivered per query (capped by the mode's `searchLimit`).
|
||
- `R` = downstream model input rate. Sonnet 4.6 = \$3/M. Opus 4.7 = \$5/M. Haiku 4.5 = \$1/M.
|
||
- `Q` = queries per month.
|
||
|
||
**Per-query input cost** (downstream agent reads the chunks):
|
||
|
||
cost_per_query = T × N × R
|
||
|
||
| Mode | T (tokens) | N (chunks) | Sonnet (\$3/M) | Opus (\$5/M) | Haiku (\$1/M) |
|
||
|---|---|---|---|---|---|
|
||
| conservative (4K cap, 10 max) | ~400 | 10 (or fewer if budget hits) | \$0.012 | \$0.020 | \$0.004 |
|
||
| balanced (12K cap, 25 max) | ~400 | ~25 | \$0.030 | \$0.050 | \$0.010 |
|
||
| tokenmax (no cap, 50 max) | ~400 | ~50 | \$0.060 | \$0.100 | \$0.020 |
|
||
|
||
**Monthly cost** (Q × per-query):
|
||
|
||
| Mode @ Sonnet | 1K Q/mo | 10K Q/mo | 100K Q/mo |
|
||
|---|---|---|---|
|
||
| conservative | \$12 | \$120 | \$1,200 |
|
||
| balanced | \$30 | \$300 | \$3,000 |
|
||
| tokenmax | \$60 | \$600 | \$6,000 |
|
||
|
||
| Mode @ Opus | 1K Q/mo | 10K Q/mo | 100K Q/mo |
|
||
|---|---|---|---|
|
||
| conservative | \$20 | \$200 | \$2,000 |
|
||
| balanced | \$50 | \$500 | \$5,000 |
|
||
| tokenmax | \$100 | \$1,000 | \$10,000 |
|
||
|
||
**gbrain's own cost** on top:
|
||
- Query embedding (text-embedding-3-large @ \$0.13/M tokens): ~\$0.00001 per query. Negligible at every scale.
|
||
- Tokenmax Haiku expansion call (\$1/M input, \$5/M output, ~500 input + 200 output per call): ~\$0.0015 per query, or \$150/mo at 100K queries. Cache hits cut this in half.
|
||
- Per-page indexing (one-time): bounded by your import volume, not query volume. Not modeled here.
|
||
|
||
**Cache hit adjustment.** A warmed brain typically sees 30-50% cache hits on repeat-query traffic. Cache hits skip the downstream input cost entirely (the cached result was already in the agent's context once). So real-world costs run ~50-70% of the table above on a busy brain.
|
||
|
||
**Why these numbers DRIFT from your actual bill:**
|
||
- Your agent's system prompt + reasoning tokens add input that gbrain doesn't see.
|
||
- Compaction reduces input over a long session.
|
||
- Most agents make 1-5 searches per turn; cost-per-turn is what bills you, not cost-per-query.
|
||
- The model price column drifts as providers reprice; pin the rate via `src/core/anthropic-pricing.ts` for a current snapshot.
|
||
|
||
The picker copy + CLAUDE.md table are the canonical user-facing source. Update them in lockstep when the underlying chunker size or default `searchLimit` changes.
|
||
|
||
## Mode × Model matrix (the 25x spread)
|
||
|
||
The per-query math above assumes Sonnet 4.6 downstream. In reality, the
|
||
downstream model tier is the BIGGER cost lever. Per-query cost at 10K
|
||
queries/month (typical single-user volume), search payload only (no cache
|
||
savings):
|
||
|
||
| Mode (search tokens) | Haiku 4.5 (\$1/M) | Sonnet 4.6 (\$3/M) | Opus 4.7 (\$5/M) |
|
||
|---|---|---|---|
|
||
| conservative (~4K) | **\$40/mo** | \$120/mo | \$200/mo |
|
||
| balanced (~10K) | \$100/mo | \$300/mo | \$500/mo |
|
||
| tokenmax (~20K) | \$200/mo | \$600/mo | **\$1,000/mo** |
|
||
|
||
Scales linearly: multiply by 10 for 100K/mo (heavy power user / multi-user
|
||
fleet); divide by 10 for 1K/mo (light usage).
|
||
|
||
**Natural pairings span ~4x** (cheap model + tight mode → frontier model + loose
|
||
mode). **Mismatches waste capacity:**
|
||
|
||
- `tokenmax + Haiku`: Haiku gets 20K of search results stuffed into its
|
||
context per query. Haiku's reasoning is weaker; more chunks = more noise,
|
||
not more signal. You pay Haiku rates but get sub-Haiku quality. Wrong
|
||
direction.
|
||
- `conservative + Opus`: Opus has 200K context window and can synthesize
|
||
across many chunks. Capping at 10 chunks / 4K tokens leaves Opus
|
||
reasoning underfed. You pay Opus rates but get conservative-shape
|
||
retrieval. Wasted spend.
|
||
|
||
**Right-sizing rule:** match the mode's `searchLimit` to the downstream
|
||
model's "useful context depth":
|
||
|
||
- Haiku struggles past ~5-10 chunks of cross-referenced content → conservative
|
||
- Sonnet handles ~25-40 chunks well → balanced
|
||
- Opus benefits from 50+ chunks for multi-hop reasoning → tokenmax
|
||
|
||
## Realistic-scale anchor (single power-user agent loop)
|
||
|
||
The per-query math above is honest but theoretical: it treats each search as an isolated billable event. Real agent loops amortize a lot of context across turns via Anthropic prompt caching. Here's what one heavy power-user loop actually looks like in production, anonymized + scaled so the numbers represent a representative power user rather than any specific deployment.
|
||
|
||
**Reference shape — tokenmax in production at a single-user scale:**
|
||
|
||
| Quantity | Approximate value |
|
||
|---|---|
|
||
| 30-day total agent spend | ~\$700/mo |
|
||
| 30-day total tokens billed | ~800M |
|
||
| Turns per month | ~860 (~29/day; one active agent loop) |
|
||
| Average tokens per turn | ~900K |
|
||
| Average cost per turn | ~\$0.85 |
|
||
| Anthropic prompt-cache hit rate | ~88% |
|
||
|
||
A "turn" here is one agent loop iteration: read user message, plan, execute tool calls (including gbrain searches), generate response. Each turn typically includes 2-4 gbrain searches.
|
||
|
||
**Per-mode scaling from the tokenmax anchor:**
|
||
|
||
The cost difference between modes is concentrated in the search-attributable fraction of per-turn cost. System prompt, tool definitions, conversation history, and reasoning tokens don't change with mode — only the chunks gbrain delivers do. Assume 3 searches per turn at the mode's `searchLimit`:
|
||
|
||
| Mode | Search tokens/turn | Search cost/turn (at \$3/M effective) | Search-attributable @ 860 turns | Δ vs tokenmax |
|
||
|---|---|---|---|---|
|
||
| tokenmax | ~60K (3 × 20K) | ~\$0.18 | ~\$155/mo | — |
|
||
| balanced | ~30K (3 × 10K) | ~\$0.09 | ~\$77/mo | -\$78 |
|
||
| conservative | ~12K (3 × 4K) | ~\$0.036 | ~\$31/mo | -\$124 |
|
||
|
||
**Implied total agent spend by NATURAL PAIRING** (mode + matched
|
||
downstream model). Per-turn cost scales with the downstream model's
|
||
per-token rate, since the cached prefix + uncached portion + reasoning
|
||
tokens all bill at that rate:
|
||
|
||
| Pairing | Per-turn cost | Total @ 860 turns/mo |
|
||
|---|---|---|
|
||
| tokenmax + Opus (frontier, max quality) | ~\$0.85 | ~\$700/mo |
|
||
| balanced + Sonnet (the sweet spot) | ~\$0.50 | ~\$430/mo |
|
||
| conservative + Haiku (cost-sensitive) | ~\$0.20 | ~\$170/mo |
|
||
|
||
**4x spread across natural pairings.** The model tier dominates because
|
||
the per-token rate applies to the WHOLE per-turn payload (system + tools
|
||
+ history + reasoning + search), not just gbrain's chunks. Mode choice
|
||
contributes ~10-20% on top of that base.
|
||
|
||
**Mismatched pairings push you off the curve:**
|
||
|
||
| Pairing | Per-turn estimate | Total @ 860 turns/mo | Compared to natural |
|
||
|---|---|---|---|
|
||
| tokenmax + Haiku | ~\$0.20 | ~\$170/mo | Same cost as conservative+Haiku, worse quality |
|
||
| conservative + Opus | ~\$0.75 | ~\$640/mo | 92% of tokenmax+Opus spend, conservative-shape retrieval |
|
||
|
||
The mismatch math says: a tokenmax+Haiku user pays the same as
|
||
conservative+Haiku but gets a noisier context (Haiku can't filter signal
|
||
from 50 chunks). A conservative+Opus user pays nearly the same as
|
||
tokenmax+Opus but starves Opus on retrieval depth. Both burn budget for
|
||
no improvement.
|
||
|
||
**What this anchor tells us that the per-query math doesn't:**
|
||
|
||
1. **At realistic agent-loop scale with disciplined prompt caching, mode choice saves 10-20% of total agent spend** — meaningful, but smaller than the per-query 5x ratio implies. Disciplined prompt-cache layouts blunt the mode delta because most of the per-turn cost is the cached prefix, not the search payload.
|
||
|
||
2. **Without that prompt-cache discipline, the per-query framing reasserts itself.** Setups that churn the prompt prefix on every turn (frequent system-prompt edits, untemplated tool defs, no prompt-cache structuring) see search payload contribute a much larger fraction of total cost. Those setups should care about mode choice more, not less.
|
||
|
||
3. **The cache hit rate quoted here (~88%) is achievable but not automatic.** It requires structuring the prompt so the cached prefix stays stable across turns: system prompt + tool defs first, history compacted but cache-aware, retrieved chunks appended LAST (where their volatility doesn't invalidate the prefix). Agents that interleave search results inside the cached region pay the prefix-rebuild tax on every turn.
|
||
|
||
**Caveats stacked here:**
|
||
|
||
- The anchor represents ONE power-user loop. Multi-user fleets aggregate proportionally; the per-user shape doesn't change.
|
||
- The "3 searches per turn" assumption varies wildly. A code-review agent might issue 10+ searches per turn; a chat-only loop might do 0.
|
||
- The 88% cache hit rate is the high end of what's achievable. Half that is closer to a default agent without cache-aware prompt layout.
|
||
- The "Δ vs tokenmax" math assumes the OTHER cost components (system, tools, history, reasoning) stay constant. In practice, conservative's smaller per-turn payload also leaves more room in the context window for history → which can change agent behavior in either direction.
|
||
|
||
This anchor + the per-query math both live in this doc on purpose. The per-query framing is what an isolated benchmark would measure (and what `gbrain eval run-all` will produce). The realistic-scale anchor is what an operator actually pays. Both are honest; neither is the whole truth.
|
||
|
||
## Reproducibility footer
|
||
|
||
Every release that publishes eval numbers includes a footer with:
|
||
|
||
- Code commit SHA
|
||
- Dataset SHA (LongMemEval, BrainBench, Replay)
|
||
- `--seed N`
|
||
- Run commands verbatim
|
||
- API model identifiers used (Anthropic + OpenAI + judge model)
|
||
|
||
Without these, the numbers are unfalsifiable. With them, anyone with API keys can re-score.
|