Performance and resource envelopes (1.0)

Status: measurement recipe + baseline notes (0.9.7); large-library UI soak pass 2026-08-07 (200+ transcripts); Medium Balanced batch pass 2026-08-07 (~9.3 min / 6 transcripts on Docker Compose); thorough full-preset LLM timings on speaker-complete corpus pass 2026-08-07 (qwen2.5:7b); final Thorough stress pass (2 tx + 2 groups) pass 2026-08-08 (corpus 68.7 min) Programme: pre_release_roadmap_1_0.md §12
Related: release_severity_triage_1_0.md, run_performance.md, runtime/docker-efficiency-baseline.md

Documented expectations and regression indicators — not necessarily strict universal guarantees. Capacity failures that corrupt data or hang without recovery are release blockers / must-fix; non-critical misses may ship as known limitations.

Representative corpus sizes

Class	Working definition	Notes
Small	1 short meeting (~2–8 minutes, low segment count)	First useful result / smoke
Medium	~5–10 transcripts; default (Balanced) preset	Typical personal corpus
Large-for-1.0	~50 transcripts in library + one multi-member group (3–5 members)	Upper documented expectation for 1.0

Raw measurement notes may live under ignored .local/ scratch; curated numbers only below.

Measurement recipe

From repo root on the machine under test (record package version, OS, Docker vs native, CPU/RAM class):

# 1) Package / image identity
python -c "import transcriptx; print(transcriptx.__version__)"
docker images --digests transcriptx:latest   # if Docker profile

# 2) Startup (GUI cold)
# Time until Home is interactive after `make run` / `./transcriptx.sh`
# Record wall seconds.

# 3) Import (Small / Medium)
# Managed import of N WhisperX/whispermlx JSON transcripts; record wall + disk delta under data root.

# 4) Default-preset runtime
# Run Balanced (or product default) on Small and Medium; read
#   <run>/.transcriptx/run_performance.json
# and module duration_ms in run_results.json (see run_performance.md).

# 5) Time to first useful result
# Clock from empty library → import Small → first Overview/Insights paint.

# 6) Memory / disk
# Peak RSS during Medium default preset; data-root disk before/after.

# 7) Model download sizes (optional extras)
# Document Hub / spaCy / Ollama pulls from runtime/models.md; do not re-download in CI.

# 8) Docker image size
docker images transcriptx:latest
docker history transcriptx:latest
# Compare docs/runtime/docker-efficiency-baseline.md (~3.7GB class historically).

# 9) Group scaling
# One group of 3–5 Medium-class members; record group wall from group run_performance sidecar.

# 10) Insufficient capacity
# Note behaviour when disk full / OOM / missing model: must fail closed with recovery path (no corrupt commit).

Optional helper (maintainer):

make perf-envelopes
# or: python3 scripts/release/perf_envelope_recipe.py

Keep notes in .local/perf_envelopes_<date>.md (gitignored). Curated numbers only below.

Metrics checklist

Metric	Status	Expectation / note
Startup time	measured-or-tagged	Target: interactive Home under ~30s cold on supported Docker/native (known limitation if host-bound)
Import time	measured-or-tagged	Small: seconds–low tens; Medium scales ~linear with file count
Time to first useful result	measured-or-tagged	Small path should complete without undocumented steps
Default-preset runtime	measured	Medium Balanced pass 2026-08-07 — 6 transcripts, batch wall ~9.3 min, all succeeded (see manual_acceptance_1_0.md §3.12). Thorough full-preset + local LLM (qwen2.5:7b) on speaker-complete corpus: see § Thorough full-preset LLM timings + § Final Thorough stress pass below.
Memory and disk use	measured-or-tagged	Record peak RSS + data-root delta; OOM without recovery = blocker
Model download sizes	documented	See runtime/models.md; first-run download is expected when enabled
Docker image size	documented baseline	Historical ~3.7GB class — docker-efficiency-baseline.md; re-measure on release hardware
Group-analysis scaling	measured-or-tagged	Group wall ≠ sum of members; includes aggregation
UI responsiveness with large library	measured	pass 2026-08-07 — Home/library responsive with 200+ transcripts (exceeds Large-for-1.0 ~50); see manual_acceptance_1_0.md §3.12
Behaviour when disk/RAM/model insufficient	must document	Fail closed; no corrupt run commit; actionable GUI/docs errors

Thorough full-preset LLM timings (2026-08-07)

Scope: every managed transcript whose speaker map is complete (every diarized ID named or ignored). After a local schema_version "1.0" → 1 patch so the files load under the integer-1 epoch, that set is:

Transcript	Segments	Speakers	Duration	Status
260615_Ana_phd_presentation_QA.json	297	10	~28 min	complete
260615_Ana_phd_supervision_meeting.json	450	3	~36 min	complete

(_deep_test_large_norm remains partial — 1 unidentified speaker — and was excluded. Mini / fixture transcripts are none / incomplete.)

Environment: native host Python 0.9.8.7, Apple Silicon / host Ollama, shared model qwen2.5:7b (~7.6B, Q4_K_M; project-recommended ~6–8B class). Mode=full, modules=None (default full set including LLM consumers + chart_descriptions finalize).

Results

Transcript	Wall	DAG	chart_descriptions finalize	Final status	Notes
Presentation QA (initial)	71.4 min (4283 s)	53.3 min	18.1 min	partial	4 LLM modules timed out at 600 s each while Ollama still held qwen3-vl:8b (~44 GB). Soft-fail continued the pipeline.
Presentation QA (LLM retry)	32 s	n/a	n/a	succeeded	Retried only llm_action_items / llm_speaker_summary / llm_summary / narrative_summary with free Ollama — all green (3–6 s each).
Supervision meeting	10.1 min (606 s)	3.9 min	6.1 min	succeeded	42 modules, 0 errors. Clean run with qwen2.5:7b resident.
Corpus effective sum	~42.0 min	—	—	composite	QA effective (~31.9 min = initial − 2400 s timeout dead time + retry) + supervision 10.1 min

LLM consumer timings (clean Ollama / qwen2.5:7b)

Module	Presentation QA (retry)	Supervision
llm_action_items	6.0 s	7.0 s
llm_speaker_summary	5.7 s	6.5 s
llm_summary	3.2 s	2.9 s
narrative_summary	5.7 s	3.0 s
topic_shift	(545 s under contention in initial QA)	13.6 s
chart_descriptions	~18 min (197-class chart set; initial QA)	~6.1 min

Regression / ops notes

• Ollama contention is a first-class risk: a resident multi-GB vision model can push LLM modules into the 600 s soft-timeout path and inflate wall clock by ~40 minutes even though non-LLM work is fine. Documented expectation: keep the intended chat model loaded (or unload others) before thorough LLM runs; treat multi-model contention hangs as capacity / ops, not analysis correctness failures when the circuit-breaker soft-fails.

• chart_descriptions dominate thorough walls once LLM chat modules are healthy (often longer than the entire DAG on chart-heavy transcripts).

• Machine-readable copies: artifacts/roadmap_1_0_llm_analysis/corpus_timings.json, qa_timings.json, plus run logs / thorough_named_speaker_corpus.md. Private mirror: .local/release_evidence/20260807_thorough_qwen25_7b/. Run dirs: …/260615_Ana_phd_presentation_QA/20260807_171909_15949872 (+ retry …/20260807_232957_38197390), …/260615_Ana_phd_supervision_meeting/20260807_233852_38732557.

Final Thorough stress pass (2026-08-08)

Scope: same speaker-complete transcript rule as above, plus every group whose members are all speaker-complete. Qualifying set on this host: 2 transcripts + 2 groups (_deep_test_bertopic_group2, Perf smoke Ana multi-speaker — identical Ana member pair).

Environment: native host Python 0.9.8.8 / git 3b206e3, Apple Silicon / host Ollama, shared model qwen2.5:7b only (no vision contention). Preset=thorough (42 suitable modules including BERTopic + LLM consumers + chart_descriptions).

Results

Run	Wall	Final status	Notes
Presentation QA	9.8 min (585.5 s)	succeeded	42 modules / 0 failures; DAG ~199 s; chart_descriptions est ~6.4 min; LLM 4/4 calls / 15.3 s logical
Supervision meeting	10.9 min (651.9 s)	succeeded	42 modules / 0 failures; DAG ~261 s; chart_descriptions est ~6.5 min; LLM 5/5 calls / 44.1 s logical
Group _deep_test_bertopic_group2	24.0 min (1442.6 s)	succeeded	both members + aggregation + group charts
Group Perf smoke Ana multi-speaker	24.0 min (1439.4 s)	succeeded	same members; independent group run
Corpus sum	68.7 min (4119.4 s)	passed	transcripts 20.6 + groups 48.0

Clean-Ollama transcript walls are ~4× faster than the contended 2026-08-07 presentation QA probe; group walls are ~2.3× the two-transcript sum (member re-analysis + aggregation + group chart_descriptions).

Artifacts

• Scratch: artifacts/roadmap_1_0_llm_analysis/stress_pass_20260808/ (SUMMARY.md, batch_summary.json, corpus_timings.json, batch.log)

• Private mirror: .local/release_evidence/20260808_thorough_stress_pass/

• Run dirs: …/260615_Ana_phd_presentation_QA/20260808_151213_94733139, …/260615_Ana_phd_supervision_meeting/20260808_152158_95318679, …/groups/7b9c6531-…/20260808_135224_e1985784, …/groups/bba6641e-…/20260808_141623_8a279f84

UI Docker Thorough batch — partial / stalled (2026-08-09)

Scope: Streamlit Batch Analysis via Docker Compose on the mounted managed library (not the native speaker-complete stress corpus above). Preset = Thorough (41 modules). Live LLM gemma3:4b. Package 0.9.8.9.

Outcome: stalled — incomplete; do not treat as a Thorough envelope pass.

Phase	Wall	Status	Notes
10 skip-heavy library-head transcripts	~2.1 min sum	succeeded	Unnamed-speaker gates → mostly SKIP; ~9–16 s each
First full Thorough transcript (tx-full-01)	~18 min	artifacts written; batch did not advance	41/41 modules, 0 failures; 197/197 chart_descriptions
Batch remainder	—	stalled	No next pipeline start after tx-full-01; .run_finalization.lock left; .transcriptx/run_performance.json missing

Ops notes: UI Thorough on a mixed library is skip-heavy until named-speaker rows; long chart_descriptions is a batch-resilience risk (logs look finished while the queue hangs). An earlier Balanced UI kick the same evening was aborted as the wrong preset.

Artifacts: artifacts/roadmap_1_0_llm_analysis/ui_thorough_batch_20260809/ (anonymised labels only) · private mirror .local/release_evidence/20260809_ui_docker_thorough_batch/

Recording

Record measured values per environment (Docker vs native) in release-evidence notes when claiming envelopes. Soft-cut for 0.9.7 allows recipe + tagged gaps; RC prefers filled Small/Medium rows on release hardware.