feat: GPU acceleration, rule-based routing with context-window safety, Granite big-window fallback + gpt-oss tuning - #124

Merged

bernardladenthin merged 68 commits into

mainfrom

claude2026_06_26_00

Jun 29, 2026

Commits on Jun 27, 2026

docs: add AI index model & prompt benchmark

bernardladenthin
and
claude
committed
docs: deepen AI index benchmark (per-file matrix, pros/cons, external corroboration)

bernardladenthin
and
claude
committed
feat: per-file prompt-cache reuse and system/user prompt split

bernardladenthin
and
claude
committed
docs: record cache_prompt optimization result (COMPARISON §10)

bernardladenthin
and
claude
committed
feat: gpt-oss reasoning config (reasoning_format=auto + reasoning_effort=low)

bernardladenthin
and
claude
committed
feat: make gpt-oss reasoning_effort configurable per model (default low)

bernardladenthin
and
claude
committed
feat: clean gpt-oss small/medium/high presets sized for large files

bernardladenthin
and
claude
committed
docs: gpt-oss reasoning_effort table + large-file clean-run validation

bernardladenthin
and
claude
committed
feat: size-tiered gpt-oss presets (c16k/c48k/c96k) for large projects

bernardladenthin
and
claude
committed
docs: large-file context limits, mixed-size strategy, 250KB validation + generator

bernardladenthin
and
claude
committed
docs: add context-size A/B (RAM-vs-time) evidence to §11

bernardladenthin
and
claude
committed
feat: per-file duration estimate (logged ETA) + gpt-oss c96k default

bernardladenthin
and
claude
committed
docs: per-file timing model (O(n^2) prefill) + c96k-default rationale in §11

bernardladenthin
and
claude
committed
docs: bring READMEs up to date with gpt-oss presets, c96k default, timing model

bernardladenthin
and
claude
committed
fix: ASCII-only per-file ETA log and "<1 KB" size label

bernardladenthin
and
claude
committed
feat: default the self-index to gpt-oss-20B-c96k

bernardladenthin
and
claude
committed
docs: production default is now gpt-oss-20B-c96k (Qwen = throughput alternative)

bernardladenthin
and
claude
committed
docs(gpt-oss-tuning): Phase 1 — reasoning_effort × budget findings

bernardladenthin
and
claude
committed
docs(gpt-oss-tuning): Phase 2 — temperature sweep findings

bernardladenthin
and
claude
committed

Commits on Jun 28, 2026

Commits on Jun 29, 2026