reeval 0.10.0 → 0.12.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +64 -39
- package/THIRD_PARTY_NOTICES.md +75 -106
- package/bin/mcp-timer.cjs +224 -0
- package/dist/{adapter-DWlpOtRn.js → adapter-DigeVFWC.js} +3 -3
- package/dist/{adapter-DWlpOtRn.js.map → adapter-DigeVFWC.js.map} +1 -1
- package/dist/{agent-CaS_2Sce.js → agent-CatUEo5h.js} +21 -6
- package/dist/{agent-CaS_2Sce.js.map → agent-CatUEo5h.js.map} +1 -1
- package/dist/{assemble-B9Alds-z.js → assemble-CF063ms6.js} +19 -17
- package/dist/assemble-CF063ms6.js.map +1 -0
- package/dist/assemble-CJoD_Rzr.js +1 -0
- package/dist/bridge/index.mjs +15 -7
- package/dist/{capture-CNb9SuU5.js → capture-BiD7lVaM.js} +26 -18
- package/dist/{capture-CNb9SuU5.js.map → capture-BiD7lVaM.js.map} +1 -1
- package/dist/capture-Dp3q7K7Y.js +1 -0
- package/dist/{cli-D303Vj5N.js → cli-B5zQCR2r.js} +266 -547
- package/dist/cli-B5zQCR2r.js.map +1 -0
- package/dist/{cli-spec-BPZbQG2p.js → cli-spec-BXIBUlaK.js} +46 -18
- package/dist/cli-spec-BXIBUlaK.js.map +1 -0
- package/dist/cli.js +2 -2
- package/dist/{command-BT9a37ZR.js → command-BMHXHqMT.js} +17 -14
- package/dist/command-BMHXHqMT.js.map +1 -0
- package/dist/command-C94CQNh8.js +91 -0
- package/dist/command-C94CQNh8.js.map +1 -0
- package/dist/command-CVtgyqlM.js +335 -0
- package/dist/command-CVtgyqlM.js.map +1 -0
- package/dist/{command-BLfT7vnQ.js → command-CaZgCTIw.js} +13 -17
- package/dist/command-CaZgCTIw.js.map +1 -0
- package/dist/{command-CMJtHKRa.js → command-DTgBZR0k.js} +6 -6
- package/dist/{command-CMJtHKRa.js.map → command-DTgBZR0k.js.map} +1 -1
- package/dist/{commands-1p-bIWhN.js → commands-BXzIDB3X.js} +1 -1
- package/dist/commands-CjpGBUWr.js +666 -0
- package/dist/commands-CjpGBUWr.js.map +1 -0
- package/dist/{config-DEgnUiU-.js → config-Dv5MG7Rp.js} +31 -34
- package/dist/config-Dv5MG7Rp.js.map +1 -0
- package/dist/{context-C1mxXh80.js → context-Dgn48BqT.js} +13 -2
- package/dist/context-Dgn48BqT.js.map +1 -0
- package/dist/contrasts-CcoMLC_y.js +101 -0
- package/dist/contrasts-CcoMLC_y.js.map +1 -0
- package/dist/db-BeqRiWLl.js +602 -0
- package/dist/db-BeqRiWLl.js.map +1 -0
- package/dist/db-CJDRVQal.js +2 -0
- package/dist/{dist-B8hJImYt.js → dist-CMjp3qqF.js} +81 -14
- package/dist/dist-CMjp3qqF.js.map +1 -0
- package/dist/{docker-sZrNrr82.js → docker-BbTlMmYi.js} +2 -2
- package/dist/{docker-sZrNrr82.js.map → docker-BbTlMmYi.js.map} +1 -1
- package/dist/{docker-7b_2SaH7.js → docker-CCXdbREk.js} +20 -4
- package/dist/{docker-7b_2SaH7.js.map → docker-CCXdbREk.js.map} +1 -1
- package/dist/{docker-yauK-9X8.js → docker-DC69dKxB.js} +1 -1
- package/dist/{doctor-BYNok4UO.js → doctor-CqttrZcs.js} +56 -22
- package/dist/doctor-CqttrZcs.js.map +1 -0
- package/dist/{egress-C4PZTSSe.js → egress-BfK_O9ce.js} +435 -31
- package/dist/egress-BfK_O9ce.js.map +1 -0
- package/dist/{environment-DNjRk0nU.js → environment-BBz1Yes9.js} +225 -12
- package/dist/environment-BBz1Yes9.js.map +1 -0
- package/dist/{evaluate-D4zeRGBF.js → evaluate-CXQ1o-pd.js} +539 -562
- package/dist/evaluate-CXQ1o-pd.js.map +1 -0
- package/dist/{evaluate-C1UVL21l.d.ts → evaluate-yjZKUexW.d.ts} +2 -2
- package/dist/facts-DtENlf9s.js +898 -0
- package/dist/facts-DtENlf9s.js.map +1 -0
- package/dist/{files-BXUdY8tr.js → files-B-eUpCZ1.js} +30 -7
- package/dist/files-B-eUpCZ1.js.map +1 -0
- package/dist/files-BsQbYz2i.js +65 -0
- package/dist/files-BsQbYz2i.js.map +1 -0
- package/dist/{files-DQmHE7pb.js → files-COHQLEHE.js} +6 -17
- package/dist/files-COHQLEHE.js.map +1 -0
- package/dist/files-DqCBW7i8.js +1 -0
- package/dist/files-Ha-jsQXf.js +1 -0
- package/dist/generate-Cs2TMH_k.js +1647 -0
- package/dist/generate-Cs2TMH_k.js.map +1 -0
- package/dist/{git-DfG-YJ1u.js → git-DhykwgKF.js} +2 -2
- package/dist/{git-DfG-YJ1u.js.map → git-DhykwgKF.js.map} +1 -1
- package/dist/{harness-B8QC9u0O.js → harness-BAjw1fcr.js} +3 -3
- package/dist/{harness-B8QC9u0O.js.map → harness-BAjw1fcr.js.map} +1 -1
- package/dist/harness.d.ts +1 -1
- package/dist/harness.js +1 -1
- package/dist/{header-hK-Z_Po9.js → header-kSmWatcO.js} +10 -3
- package/dist/{header-hK-Z_Po9.js.map → header-kSmWatcO.js.map} +1 -1
- package/dist/{images-0On0cztu.js → images-BmJRuTpm.js} +61 -13
- package/dist/images-BmJRuTpm.js.map +1 -0
- package/dist/index.d.ts +3 -3
- package/dist/index.js +4 -4
- package/dist/init-Bwc9byQr.js +543 -0
- package/dist/init-Bwc9byQr.js.map +1 -0
- package/dist/init-D7Y1ny1z.js +1 -0
- package/dist/insights-B2Xmhz9o.js +64 -0
- package/dist/insights-B2Xmhz9o.js.map +1 -0
- package/dist/{invocations-DbGfBAA_.js → invocations-CQTbP092.js} +3 -3
- package/dist/{invocations-DbGfBAA_.js.map → invocations-CQTbP092.js.map} +1 -1
- package/dist/{judge-DW_GDOmB.js → judge-Dt6KW6_O.js} +5 -3
- package/dist/judge-Dt6KW6_O.js.map +1 -0
- package/dist/judge.d.ts +7 -6
- package/dist/judge.js +8 -7
- package/dist/judge.js.map +1 -1
- package/dist/{live-jlfszAbb.js → live-DYd2QEyR.js} +2 -2
- package/dist/{live-jlfszAbb.js.map → live-DYd2QEyR.js.map} +1 -1
- package/dist/{model-h_cDtcRr.js → model-B-mel-Gv.js} +46 -20
- package/dist/model-B-mel-Gv.js.map +1 -0
- package/dist/model-BGKYVpoj.js +1 -0
- package/dist/{model-DaVCY3rb.js → model-C7tSwT7f.js} +2 -2
- package/dist/{model-DaVCY3rb.js.map → model-C7tSwT7f.js.map} +1 -1
- package/dist/{model-oqcDplJG.js → model-CWlAqmbU.js} +150 -46
- package/dist/model-CWlAqmbU.js.map +1 -0
- package/dist/model-DHY-dfoO.js +1 -0
- package/dist/{node-B3ydHo6t.js → node-D0mtjRAO.js} +3 -3
- package/dist/{node-B3ydHo6t.js.map → node-D0mtjRAO.js.map} +1 -1
- package/dist/{options-Bh2W63dK.js → options-CPUW97_7.js} +9 -12
- package/dist/options-CPUW97_7.js.map +1 -0
- package/dist/{pairs-CnInRu-G.js → pairs-DP6XOJby.js} +2 -2
- package/dist/{pairs-CnInRu-G.js.map → pairs-DP6XOJby.js.map} +1 -1
- package/dist/{patch-Zac5aIRM.js → patch-D4oluWzH.js} +2 -2
- package/dist/{patch-Zac5aIRM.js.map → patch-D4oluWzH.js.map} +1 -1
- package/dist/{plan-C9Cy27Ay.js → plan-OkwUcJQ9.js} +31 -12
- package/dist/plan-OkwUcJQ9.js.map +1 -0
- package/dist/plugins-CUWMnSNb.js +654 -0
- package/dist/plugins-CUWMnSNb.js.map +1 -0
- package/dist/{preflight-CPWpQeBI.js → preflight-B6pY89bF.js} +1 -1
- package/dist/{preflight-D7iFPKeK.js → preflight-Bjm7_ylE.js} +14 -10
- package/dist/preflight-Bjm7_ylE.js.map +1 -0
- package/dist/{prepare-doGxaAGR.js → prepare-CQmK3SBn.js} +71 -24
- package/dist/prepare-CQmK3SBn.js.map +1 -0
- package/dist/prepare-DbdgtZKe.js +2 -0
- package/dist/{provenance-DmE9uPVt.js → provenance-B3Y17oCC.js} +2 -2
- package/dist/{provenance-DmE9uPVt.js.map → provenance-B3Y17oCC.js.map} +1 -1
- package/dist/providers-CXN2RFz-.js +1 -0
- package/dist/{providers-tAOD1NQN.js → providers-eoniqK6w.js} +24 -10
- package/dist/providers-eoniqK6w.js.map +1 -0
- package/dist/query-CLRZmWBT.js +139 -0
- package/dist/query-CLRZmWBT.js.map +1 -0
- package/dist/query-CmBDYt0_.js +1 -0
- package/dist/{questions-DjAwMxZ3.js → questions-1XOulevV.js} +6 -6
- package/dist/{questions-DjAwMxZ3.js.map → questions-1XOulevV.js.map} +1 -1
- package/dist/{record-BKmMvZ6O.js → record-kk9M1hj-.js} +8 -4
- package/dist/record-kk9M1hj-.js.map +1 -0
- package/dist/remarks-1521zwfW.js +38 -0
- package/dist/remarks-1521zwfW.js.map +1 -0
- package/dist/remarks-BhAV1H4V.js +1 -0
- package/dist/remarks-DlJCqWRs.js +278 -0
- package/dist/remarks-DlJCqWRs.js.map +1 -0
- package/dist/{remote-HDRt16ew.js → remote-cE2VWfkJ.js} +3 -3
- package/dist/{remote-HDRt16ew.js.map → remote-cE2VWfkJ.js.map} +1 -1
- package/dist/resolve-C4zeccCH.js +724 -0
- package/dist/resolve-C4zeccCH.js.map +1 -0
- package/dist/resolve-D30yZYmB.js +1 -0
- package/dist/{results-DuauQI3Z.js → results-Drtt0uFl.js} +33 -7
- package/dist/results-Drtt0uFl.js.map +1 -0
- package/dist/{resume-BRfa8E_s.js → resume-BEFazUfy.js} +4 -4
- package/dist/{resume-BRfa8E_s.js.map → resume-BEFazUfy.js.map} +1 -1
- package/dist/rules-BFzQqPIm.js +317 -0
- package/dist/rules-BFzQqPIm.js.map +1 -0
- package/dist/run-BeMy92Us.js +88 -0
- package/dist/run-BeMy92Us.js.map +1 -0
- package/dist/{sandbox-D3A8oUb7.js → sandbox-CQaJj7n9.js} +2 -2
- package/dist/{sandbox-D3A8oUb7.js.map → sandbox-CQaJj7n9.js.map} +1 -1
- package/dist/sandbox.d.ts +1 -1
- package/dist/sandbox.js +1 -1
- package/dist/server-V4-2w0l8.js +1 -0
- package/dist/{server-XDleKaF_.js → server-xK0les4f.js} +8 -8
- package/dist/{server-XDleKaF_.js.map → server-xK0les4f.js.map} +1 -1
- package/dist/{shared-DYpIb1BI.js → shared-CA61BuPJ.js} +25 -3
- package/dist/shared-CA61BuPJ.js.map +1 -0
- package/dist/source-BDow_oW8.js +1 -0
- package/dist/{source-C8xumIWj.js → source-Q1hEKbCs.js} +2 -2
- package/dist/{source-C8xumIWj.js.map → source-Q1hEKbCs.js.map} +1 -1
- package/dist/src-B9L9yGKI.js +1 -0
- package/dist/{src-ofUaYbuY.js → src-Bcah6ZkT.js} +1 -1
- package/dist/{src-DvCg0NBs.js → src-BhmaorOU.js} +28 -6
- package/dist/{src-DvCg0NBs.js.map → src-BhmaorOU.js.map} +1 -1
- package/dist/{src-CF17Ug08.js → src-C0CeKiLU.js} +19 -4
- package/dist/{src-CF17Ug08.js.map → src-C0CeKiLU.js.map} +1 -1
- package/dist/src-C0xowZ3C.js +2 -0
- package/dist/src-CHIfIKUu.js +2 -0
- package/dist/src-DLtyz4w5.js +2 -0
- package/dist/src-DPIjmReZ.js +2 -0
- package/dist/{src-C_lPtowT.js → src-DXNh5mTr.js} +1 -1
- package/dist/{src-FGXhkEkH.js → src-DiLMkIr9.js} +142 -13
- package/dist/src-DiLMkIr9.js.map +1 -0
- package/dist/{src-B7treWa8.js → src-LekQrr_S.js} +324 -32
- package/dist/src-LekQrr_S.js.map +1 -0
- package/dist/src-bCnsA6Vp.js +2 -0
- package/dist/{src-u8_ixn_F.js → src-cbCb-MPy.js} +20 -4
- package/dist/src-cbCb-MPy.js.map +1 -0
- package/dist/{src-DvK0EIZx.js → src-qGoFkuJF.js} +55 -6
- package/dist/{src-DvK0EIZx.js.map → src-qGoFkuJF.js.map} +1 -1
- package/dist/{src-jzagbuqH.js → src-ynBVosfX.js} +1 -1
- package/dist/state-BQR_BCm4.js +1 -0
- package/dist/{state-C__9swoG.js → state-BmZ8QVZ8.js} +12 -12
- package/dist/{state-C__9swoG.js.map → state-BmZ8QVZ8.js.map} +1 -1
- package/dist/studio-app/assets/index-CnYJijyI.js +58 -0
- package/dist/studio-app/index.html +1 -1
- package/dist/suite-BaHPmAGy.js +1 -0
- package/dist/suite-C8sw2pox.js +481 -0
- package/dist/suite-C8sw2pox.js.map +1 -0
- package/dist/summary-0oQNGskv.js +837 -0
- package/dist/summary-0oQNGskv.js.map +1 -0
- package/dist/testing.d.ts +2 -2
- package/dist/testing.js +2 -2
- package/dist/{text-gEcKjAyp.js → text-CrVy5ioL.js} +18 -2
- package/dist/text-CrVy5ioL.js.map +1 -0
- package/dist/{transcript-BNNANU8A.js → transcript-BTxJpW4R.js} +2 -2
- package/dist/{transcript-BNNANU8A.js.map → transcript-BTxJpW4R.js.map} +1 -1
- package/dist/types-C5vI3g1C.js.map +1 -1
- package/dist/{types-g17bRAzt.d.ts → types-GraZtfHN.d.ts} +858 -39
- package/dist/usage-BDo2q_yu.js +1 -0
- package/dist/{command-Bge8F7Ko.js → validate-BdoXUuL_.js} +35 -383
- package/dist/validate-BdoXUuL_.js.map +1 -0
- package/dist/view/index.d.ts +20 -9
- package/dist/view/index.js +7 -6
- package/dist/views-BoUoX4_A.js +125 -0
- package/dist/views-BoUoX4_A.js.map +1 -0
- package/dist/views-noVQNCUP.js +2 -0
- package/dist/{worker-entry-yOchGYYv.js → worker-entry-D0-GErmU.js} +2 -2
- package/dist/{worker-entry-yOchGYYv.js.map → worker-entry-D0-GErmU.js.map} +1 -1
- package/dist/{workspace-D8sSgpiy.js → workspace-B90YDU_a.js} +1 -1
- package/dist/{workspace-CgyHEM0l.js → workspace-tjZFkjxC.js} +4 -4
- package/dist/{workspace-CgyHEM0l.js.map → workspace-tjZFkjxC.js.map} +1 -1
- package/package.json +10 -8
- package/skills/eval-rubric/SKILL.md +130 -0
- package/dist/agent-CURAwx_q.js +0 -334
- package/dist/agent-CURAwx_q.js.map +0 -1
- package/dist/assemble-B9Alds-z.js.map +0 -1
- package/dist/assemble-J5MukZFY.js +0 -1
- package/dist/capture-jquXZBKK.js +0 -1
- package/dist/cli-D303Vj5N.js.map +0 -1
- package/dist/cli-spec-BPZbQG2p.js.map +0 -1
- package/dist/command-BLfT7vnQ.js.map +0 -1
- package/dist/command-BT9a37ZR.js.map +0 -1
- package/dist/command-Bge8F7Ko.js.map +0 -1
- package/dist/command-BkFjijN2.js +0 -1
- package/dist/commands-BFx16WB8.js +0 -225
- package/dist/commands-BFx16WB8.js.map +0 -1
- package/dist/config-DEgnUiU-.js.map +0 -1
- package/dist/context-C1mxXh80.js.map +0 -1
- package/dist/db-CdHizc4_.js +0 -1406
- package/dist/db-CdHizc4_.js.map +0 -1
- package/dist/db-Dm1cRJJP.js +0 -2
- package/dist/dist-B8hJImYt.js.map +0 -1
- package/dist/doctor-BYNok4UO.js.map +0 -1
- package/dist/duration-CEOQww97.js +0 -26
- package/dist/duration-CEOQww97.js.map +0 -1
- package/dist/egress-C4PZTSSe.js.map +0 -1
- package/dist/environment-DNjRk0nU.js.map +0 -1
- package/dist/evaluate-D4zeRGBF.js.map +0 -1
- package/dist/family-D2b7WD10.js +0 -67
- package/dist/family-D2b7WD10.js.map +0 -1
- package/dist/files-BXUdY8tr.js.map +0 -1
- package/dist/files-BhyltlDB.js +0 -1
- package/dist/files-DQmHE7pb.js.map +0 -1
- package/dist/files-ZZPqfAo6.js +0 -1
- package/dist/images-0On0cztu.js.map +0 -1
- package/dist/init-B77Rpcqc.js +0 -360
- package/dist/init-B77Rpcqc.js.map +0 -1
- package/dist/init-BrOWin3W.js +0 -1
- package/dist/judge-DW_GDOmB.js.map +0 -1
- package/dist/model-NM3ID0uR.js +0 -1
- package/dist/model-h_cDtcRr.js.map +0 -1
- package/dist/model-nwWJvaUR.js +0 -1
- package/dist/model-oqcDplJG.js.map +0 -1
- package/dist/options-Bh2W63dK.js.map +0 -1
- package/dist/plan-C9Cy27Ay.js.map +0 -1
- package/dist/preflight-D7iFPKeK.js.map +0 -1
- package/dist/prepare-BGjePFQl.js +0 -2
- package/dist/prepare-doGxaAGR.js.map +0 -1
- package/dist/process-CSVO_jsc.js +0 -119
- package/dist/process-CSVO_jsc.js.map +0 -1
- package/dist/providers-gDMko8fl.js +0 -1
- package/dist/providers-tAOD1NQN.js.map +0 -1
- package/dist/query-Bes9Kh6Y.js +0 -91
- package/dist/query-Bes9Kh6Y.js.map +0 -1
- package/dist/query-C_G_svLU.js +0 -1
- package/dist/record-BKmMvZ6O.js.map +0 -1
- package/dist/results-DuauQI3Z.js.map +0 -1
- package/dist/server-C7sg1ndo.js +0 -1
- package/dist/shared-DYpIb1BI.js.map +0 -1
- package/dist/source-fGd0SaKP.js +0 -1
- package/dist/src-B7treWa8.js.map +0 -1
- package/dist/src-DNnG4e4R.js +0 -2
- package/dist/src-DTbrY1hi.js +0 -2
- package/dist/src-DdyJEh2v.js +0 -2
- package/dist/src-Dj9PAroc.js +0 -2
- package/dist/src-Du8H85Qq.js +0 -2
- package/dist/src-FGXhkEkH.js.map +0 -1
- package/dist/src-pJZOrYR9.js +0 -1
- package/dist/src-u8_ixn_F.js.map +0 -1
- package/dist/state-CgjQo29K.js +0 -1
- package/dist/studio-app/assets/index-DrQc2Yxd.js +0 -58
- package/dist/summary-BLAVLySU.js +0 -385
- package/dist/summary-BLAVLySU.js.map +0 -1
- package/dist/summary-DgyjUh_l.js +0 -192
- package/dist/summary-DgyjUh_l.js.map +0 -1
- package/dist/text-gEcKjAyp.js.map +0 -1
- package/dist/tools-GCyNYSc1.js +0 -249
- package/dist/tools-GCyNYSc1.js.map +0 -1
- package/dist/views-BW6oCDUQ.js +0 -2
- package/dist/views-BmQF4UCI.js +0 -225
- package/dist/views-BmQF4UCI.js.map +0 -1
|
@@ -65,6 +65,8 @@ async function withRetries(call, policy, signal) {
|
|
|
65
65
|
//#region src/judge/index.ts
|
|
66
66
|
/** The judge the docs recommend, pinned so a model update never silently moves scores. */
|
|
67
67
|
const RECOMMENDED_JUDGE_MODEL = "typesafe-ai/jev-1.14.0";
|
|
68
|
+
/** The language model `reeval init` sets as `analysis.model` and the docs recommend for insights. */
|
|
69
|
+
const RECOMMENDED_ANALYSIS_MODEL = "google/gemini-3.8-flash";
|
|
68
70
|
const DEFAULT_THRESHOLD = .8;
|
|
69
71
|
const GATEWAY_KEYS = ["AI_GATEWAY_API_KEY", "VERCEL_OIDC_TOKEN"];
|
|
70
72
|
/** Providers reeval calls directly, through their own AI SDK package, when their key is set. */
|
|
@@ -128,7 +130,7 @@ async function runJudge(question, state, options, signal, retry = JUDGE_RETRY) {
|
|
|
128
130
|
const problem = checkModelString(model);
|
|
129
131
|
if (problem) throw new Error(problem);
|
|
130
132
|
}
|
|
131
|
-
const { scoreWithModel } = await import("./model-
|
|
133
|
+
const { scoreWithModel } = await import("./model-C7tSwT7f.js");
|
|
132
134
|
const scored = await withRetries(() => scoreWithModel(model, question, state, options.providerOptions, signal), retry, signal);
|
|
133
135
|
({score, resolvedModel} = scored.value);
|
|
134
136
|
retries = scored.retries;
|
|
@@ -146,6 +148,6 @@ async function runJudge(question, state, options, signal, retry = JUDGE_RETRY) {
|
|
|
146
148
|
};
|
|
147
149
|
}
|
|
148
150
|
//#endregion
|
|
149
|
-
export {
|
|
151
|
+
export { checkModelString as a, judgeModelName as c, isApiError as d, RECOMMENDED_JUDGE_MODEL as i, judgeRoute as l, DIRECT_PROVIDERS as n, isFakeJudge as o, RECOMMENDED_ANALYSIS_MODEL as r, judgeCredentials as s, DEFAULT_THRESHOLD as t, runJudge as u };
|
|
150
152
|
|
|
151
|
-
//# sourceMappingURL=judge-
|
|
153
|
+
//# sourceMappingURL=judge-Dt6KW6_O.js.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"judge-Dt6KW6_O.js","names":["sleep"],"sources":["../src/judge/retry.ts","../src/judge/index.ts"],"mappings":";;;AAOA,MAAa,cAA2B;CAAE,UAAU;CAAG,QAAQ;CAAO,OAAO;AAAO;AAEpF,MAAM,mCAAmB,IAAI,IAAI;CAAC;CAAK;CAAK;CAAK;AAAG,CAAC;AACrD,MAAM,oBACJ;;AAUF,MAAM,aAAa,UAA4B;CAC7C,MAAM,IAAI;CACV,OAAO,GAAG,SAAS,mBAAmB,EAAE,YAAY,UAAU,EAAE,SAAS,IAAI;AAC/E;;AAGA,MAAa,cAAc,UACzB,OAAQ,UAAU,KAAK,CAAC,EAAoB,eAAe;;AAG7D,SAAgB,YAAY,OAAyB;CACnD,MAAM,IAAI,UAAU,KAAK;CACzB,IAAI,CAAC,KAAK,EAAE,SAAS,gBAAgB,EAAE,SAAS,gBAAgB,OAAO;CACvE,IAAI,OAAO,EAAE,eAAe,UAC1B,OAAO,EAAE,cAAc,OAAO,iBAAiB,IAAI,EAAE,UAAU,KAAK,EAAE,gBAAgB;CAExF,OAAO,EAAE,gBAAgB,QAAQ,kBAAkB,KAAK,EAAE,WAAW,OAAO,CAAC,CAAC;AAChF;AAEA,SAAS,aAAa,OAAwB;CAC5C,MAAM,SAAU,UAAU,KAAK,CAAC,EAAoB,kBAAkB;CACtE,MAAM,UAAU,OAAO,MAAM;CAC7B,OAAO,OAAO,SAAS,OAAO,KAAK,UAAU,IAAI,UAAU,MAAO;AACpE;;AAGA,SAAgB,UAAU,QAAqB,OAAe,OAAwB;CACpF,MAAM,MAAM,KAAK,IAAI,OAAO,OAAO,OAAO,SAAS,KAAK,KAAK;CAC7D,MAAM,WAAW,MAAM,IAAI,KAAK,OAAO,KAAK,MAAM;CAClD,OAAO,KAAK,MAAM,KAAK,IAAI,OAAO,OAAO,KAAK,IAAI,UAAU,aAAa,KAAK,CAAC,CAAC,CAAC;AACnF;;AAKA,eAAsB,YACpB,MACA,QACA,QAC8C;CAC9C,MAAM,UAAwB,CAAC;CAC/B,SACE,IAAI;EACF,OAAO;GAAE,OAAO,MAAM,KAAK;GAAG;EAAQ;CACxC,SAAS,OAAO;EACd,IAAI,OAAO,WAAW,QAAQ,SAAS,KAAK,OAAO,YAAY,CAAC,YAAY,KAAK,GAAG;GAClF,MAAM,SAAU,iBAAiB,QAAQ,QAAQ,IAAI,MAAM,OAAO,KAAK,CAAC;GACxE,OAAO,UAAU;GACjB,MAAM;EACR;EACA,MAAM,UAAU,UAAU,QAAQ,QAAQ,QAAQ,KAAK;EACvD,MAAM,IAAI,UAAU,KAAK;EACzB,QAAQ,KAAK;GAAE,OAAO,GAAG,WAAW,OAAO,CAAC;GAAG;EAAQ,CAAC;EACxD,MAAMA,WAAM,SAAS,KAAA,GAAW,EAAE,OAAO,CAAC;CAC5C;AAEJ;;;;ACxEA,MAAa,0BAA0B;;AAEvC,MAAa,6BAA6B;AAC1C,MAAa,oBAAoB;AAEjC,MAAM,eAAe,CAAC,sBAAsB,mBAAmB;;AAG/D,MAAa,mBAAiE;CAC5E,QAAQ;EAAE,KAAK;EAAgC,KAAK;CAAiB;CACrE,eAAe;EAAE,KAAK;EAAuB,KAAK;CAAsB;AAC1E;AAMA,MAAa,eAAe,MAC1B,OAAO,MAAM,YAAY,MAAM,QAAS,EAAqB,SAAS;AAExE,SAAgB,eAAe,OAA2B;CACxD,IAAI,OAAO,UAAU,UAAU,OAAO;CACtC,IAAI,YAAY,KAAK,GAAG,OAAO;CAC/B,MAAM,IAAI;CACV,OAAO,EAAE,YAAY,EAAE,UAAU,GAAG,EAAE,SAAS,GAAG,EAAE,YAAY;AAClE;;;;;AAMA,SAAgB,iBACd,OACA,OAAO,mBACP,WAAW,CAAC,yBAAyB,UAAU,OAAO,GAClC;CACpB,IAAI,MAAM,SAAS,GAAG,GAAG,OAAO,KAAA;CAChC,OAAO,GAAG,KAAK,iCAAiC,SAAS,KAAK,MAAM,IAAI,EAAE,EAAE,CAAC,CAAC,KAAK,MAAM,EAAE,SAAS,MAAM;AAC5G;AAEA,MAAM,cAAc,UAAkB,MAAM,MAAM,GAAG,MAAM,QAAQ,GAAG,CAAC;;AAGvE,SAAgB,iBAAiB,OAA0C;CACzE,IAAI,YAAY,KAAK,GAAG,OAAO,CAAC;CAChC,MAAM,SAAS,iBAAiB,WAAW,KAAK;CAChD,OAAO,SAAS,CAAC,OAAO,KAAK,GAAG,YAAY,IAAI;AAClD;;;;;AAMA,SAAgB,WAAW,OAAe,MAAyB,QAAQ,KAAiB;CAC1F,MAAM,WAAW,WAAW,KAAK;CACjC,MAAM,SAAS,iBAAiB;CAChC,IAAI,UAAU,IAAI,OAAO,MAAM,OAAO;EAAE,MAAM;EAAU;EAAU,KAAK,OAAO;CAAI;CAClF,IAAI,aAAa,MAAM,MAAM,IAAI,EAAE,GAAG,OAAO,EAAE,MAAM,UAAU;CAC/D,MAAM,UAAU,SACZ,OAAO,OAAO,IAAI,WAAW,OAAO,IAAI,kIACxC;CACJ,MAAM,IAAI,MAAM,mCAAmC,MAAM,KAAK,SAAS;AACzE;;AAMA,eAAsB,SACpB,UACA,OACA,SACA,QACA,QAAqB,aACC;CACtB,MAAM,EAAE,OAAO,cAAc;CAC7B,IAAI;CACJ,IAAI;CACJ,IAAI,UAAwB,CAAC;CAC7B,IAAI,YAAY,KAAK,GACnB,QAAQ,OAAO,MAAM,UAAU,aAAa,MAAM,MAAM,KAAK,IAAI,MAAM;MAClE;EACL,IAAI,OAAO,UAAU,UAAU;GAC7B,MAAM,UAAU,iBAAiB,KAAK;GACtC,IAAI,SAAS,MAAM,IAAI,MAAM,OAAO;EACtC;EACA,MAAM,EAAE,mBAAmB,MAAM,OAAO;EACxC,MAAM,SAAS,MAAM,kBACb,eAAe,OAAO,UAAU,OAAO,QAAQ,iBAAiB,MAAM,GAC5E,OACA,MACF;EACC,CAAC,CAAE,OAAO,iBAAkB,OAAO;EACpC,UAAU,OAAO;CACnB;CASA,OAAO;EAAE,SAAA;GAPP,SAAS,SAAS,YAAY,SAAS;GACvC;GACA;GACA;GACA,OAAO,eAAe,KAAK;GAC3B,GAAI,gBAAgB,EAAE,cAAc,IAAI,CAAC;EAE5B;EAAG;CAAQ;AAC5B"}
|
package/dist/judge.d.ts
CHANGED
|
@@ -1,4 +1,4 @@
|
|
|
1
|
-
import { D as ConfigJudgeOptions,
|
|
1
|
+
import { Bt as Verdict, D as ConfigJudgeOptions, J as FakeJudgeModel, K as FakeInsights, W as FakeAnalysisModel, Z as FakeVote, ct as JudgeArtifactRef, dt as JudgeModel, ft as JudgeOptions, lt as JudgeCallOptions, n as AnalysisConfig, o as BlindPair, pt as JudgeState, q as FakeInsightsInput, ut as JudgeFailure } from "./types-GraZtfHN.js";
|
|
2
2
|
//#region src/judge/index.d.ts
|
|
3
3
|
/** The judge the docs recommend, pinned so a model update never silently moves scores. */
|
|
4
4
|
export declare const RECOMMENDED_JUDGE_MODEL = "typesafe-ai/jev-1.14.0";
|
|
@@ -16,11 +16,12 @@ export declare const fakeJudge: ({ score }?: {
|
|
|
16
16
|
* An analysis model for tests: `reeval analyze` gets `output` (or `output(digest)`) back as the
|
|
17
17
|
* model's structured answer, and each compare question's pair gets `vote` (a fixed `'A'`, `'B'` or
|
|
18
18
|
* `'tie'`, `{ winner, reason? }`, a function of the blinded pair, or `{ error }`), through the same
|
|
19
|
-
* parsing and validation as a real one.
|
|
20
|
-
*
|
|
21
|
-
*
|
|
19
|
+
* parsing and validation as a real one. Insights (in `reeval run` and `reeval analyze --insights`)
|
|
20
|
+
* get `insights` (or `insights(input)`, called per eval for its classification and its sentences,
|
|
21
|
+
* then once for the headline), through the same validator: an uncited sentence is dropped, and one
|
|
22
|
+
* with a banned word is asked for again, then dropped.
|
|
22
23
|
*/
|
|
23
|
-
export declare const fakeAnalysis: ({ output,
|
|
24
|
+
export declare const fakeAnalysis: ({ output, vote, insights }: Pick<FakeAnalysisModel, 'output' | 'vote' | 'insights'>) => FakeAnalysisModel;
|
|
24
25
|
//#endregion
|
|
25
|
-
export type { AnalysisConfig, BlindPair, ConfigJudgeOptions, FakeAnalysisModel,
|
|
26
|
+
export type { AnalysisConfig, BlindPair, ConfigJudgeOptions, FakeAnalysisModel, FakeInsights, FakeInsightsInput, FakeJudgeModel, FakeVote, JudgeArtifactRef, JudgeCallOptions, JudgeFailure, JudgeModel, JudgeOptions, JudgeState, Verdict };
|
|
26
27
|
//# sourceMappingURL=judge.d.ts.map
|
package/dist/judge.js
CHANGED
|
@@ -1,4 +1,4 @@
|
|
|
1
|
-
import {
|
|
1
|
+
import { i as RECOMMENDED_JUDGE_MODEL, t as DEFAULT_THRESHOLD } from "./judge-Dt6KW6_O.js";
|
|
2
2
|
//#region src/judge.ts
|
|
3
3
|
/**
|
|
4
4
|
* A judge that returns `score` without calling a model, for tests and dry runs. `score` can be a
|
|
@@ -12,15 +12,16 @@ const fakeJudge = ({ score = 1 } = {}) => ({
|
|
|
12
12
|
* An analysis model for tests: `reeval analyze` gets `output` (or `output(digest)`) back as the
|
|
13
13
|
* model's structured answer, and each compare question's pair gets `vote` (a fixed `'A'`, `'B'` or
|
|
14
14
|
* `'tie'`, `{ winner, reason? }`, a function of the blinded pair, or `{ error }`), through the same
|
|
15
|
-
* parsing and validation as a real one.
|
|
16
|
-
*
|
|
17
|
-
*
|
|
15
|
+
* parsing and validation as a real one. Insights (in `reeval run` and `reeval analyze --insights`)
|
|
16
|
+
* get `insights` (or `insights(input)`, called per eval for its classification and its sentences,
|
|
17
|
+
* then once for the headline), through the same validator: an uncited sentence is dropped, and one
|
|
18
|
+
* with a banned word is asked for again, then dropped.
|
|
18
19
|
*/
|
|
19
|
-
const fakeAnalysis = ({ output,
|
|
20
|
+
const fakeAnalysis = ({ output, vote, insights }) => ({
|
|
20
21
|
kind: "fake-analysis",
|
|
21
22
|
...output !== void 0 ? { output } : {},
|
|
22
|
-
...
|
|
23
|
-
...
|
|
23
|
+
...vote !== void 0 ? { vote } : {},
|
|
24
|
+
...insights !== void 0 ? { insights } : {}
|
|
24
25
|
});
|
|
25
26
|
//#endregion
|
|
26
27
|
export { DEFAULT_THRESHOLD, RECOMMENDED_JUDGE_MODEL, fakeAnalysis, fakeJudge };
|
package/dist/judge.js.map
CHANGED
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"judge.js","names":[],"sources":["../src/judge.ts"],"mappings":";;;;;;AAQA,MAAa,aAAa,EACxB,QAAQ,MAGN,CAAC,OAAuB;CAC1B,MAAM;CACN;AACF
|
|
1
|
+
{"version":3,"file":"judge.js","names":[],"sources":["../src/judge.ts"],"mappings":";;;;;;AAQA,MAAa,aAAa,EACxB,QAAQ,MAGN,CAAC,OAAuB;CAC1B,MAAM;CACN;AACF;;;;;;;;;;AAWA,MAAa,gBAAgB,EAC3B,QACA,MACA,gBACiF;CACjF,MAAM;CACN,GAAI,WAAW,KAAA,IAAY,EAAE,OAAO,IAAI,CAAC;CACzC,GAAI,SAAS,KAAA,IAAY,EAAE,KAAK,IAAI,CAAC;CACrC,GAAI,aAAa,KAAA,IAAY,EAAE,SAAS,IAAI,CAAC;AAC/C"}
|
|
@@ -1,5 +1,5 @@
|
|
|
1
1
|
import { n as isOneOf } from "./object-CLf2-tNq.js";
|
|
2
|
-
import { n as fillDefaults } from "./record-
|
|
2
|
+
import { n as fillDefaults } from "./record-kk9M1hj-.js";
|
|
3
3
|
//#region src/view/live.ts
|
|
4
4
|
/** The progress file's name inside the run's staging directory. */
|
|
5
5
|
const LIVE_FILE = "live.json";
|
|
@@ -84,4 +84,4 @@ function readableLiveProgress(value) {
|
|
|
84
84
|
//#endregion
|
|
85
85
|
export { liveApiPath as a, readableLiveProgress as c, isFinished as i, LIVE_FILE as n, liveCounts as o, LIVE_SCHEMA_VERSION as r, liveTranscriptPath as s, FINISHED as t };
|
|
86
86
|
|
|
87
|
-
//# sourceMappingURL=live-
|
|
87
|
+
//# sourceMappingURL=live-DYd2QEyR.js.map
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"live-
|
|
1
|
+
{"version":3,"file":"live-DYd2QEyR.js","names":[],"sources":["../src/view/live.ts"],"mappings":";;;;AAQA,MAAa,YAAY;AACzB,MAAa,sBAAsB;AAsDnC,MAAa,cAAc;AAC3B,MAAa,sBAAsB,OAAe,eAChD,YAAY,MAAM,sBAAsB,mBAAmB,UAAU;AAEvE,MAAa,WAAkC;CAAC;CAAQ;CAAQ;CAAe;AAAS;AACxF,MAAa,cAAc,WAAuB,SAAS,SAAS,MAAM;AAE1E,SAAgB,WAAW,UAA8D;CACvF,MAAM,SAAS,GAAG,aAChB,SAAS,QAAQ,MAAM,SAAS,SAAS,EAAE,MAAM,CAAC,CAAC,CAAC;CACtD,OAAO;EACL,MAAM,MAAM,GAAG,QAAQ;EACvB,QAAQ,MAAM,MAAM;EACpB,QAAQ,MAAM,MAAM;EACpB,SAAS,MAAM,aAAa;EAC5B,SAAS,MAAM,SAAS;EACxB,SAAS,MAAM,SAAS;CAC1B;AACF;AAEA,MAAM,WAAkC;CAAC;CAAU;CAAW,GAAG;AAAQ;AACzE,MAAM,SAA+B;CAAC;CAAW;CAAc;AAAW;AAC1E,MAAM,SAAS,MACb,OAAO,MAAM,YAAY,MAAM,QAAQ,CAAC,MAAM,QAAQ,CAAC;AACzD,MAAM,OAAO,MAA4B,OAAO,MAAM;AACtD,MAAM,OAAO,MAA4B,OAAO,MAAM,YAAY,OAAO,SAAS,CAAC;AACnF,MAAM,UAAa,GAAY,OAC7B,MAAM,QAAQ,MAAM,KAAA,KAAa,GAAG,CAAC;AAEvC,SAAS,gBAAgB,GAAgC;CACvD,IAAI,CAAC,MAAM,CAAC,GAAG,OAAO;CACtB,MAAM,EAAE,KAAK,QAAQ,MAAM,MAAM,MAAM,GAAG,QAAQ,WAAW,eAAe;CAC5E,IAAI,CAAC,IAAI,GAAG,KAAK,CAAC,IAAI,MAAM,KAAK,CAAC,IAAI,IAAI,KAAK,CAAC,IAAI,IAAI,KAAK,CAAC,IAAI,IAAI,KAAK,CAAC,IAAI,CAAC,GAAG,OAAO;CAC3F,IAAI,CAAC,QAAQ,UAAU,MAAM,GAAG,OAAO;CACvC,IAAI,CAAC,OAAO,WAAW,GAAG,KAAK,CAAC,OAAO,YAAY,GAAG,GAAG,OAAO;CAChE,OAAO;EACL;EACA;EACA;EACA;EACA;EACA;EACA;EACA,WAAW,aAAa;EACxB,YAAY,cAAc;CAC5B;AACF;;;;;AAMA,SAAgB,qBAAqB,OAAqC;CACxE,IAAI,CAAC,MAAM,KAAK,KAAK,MAAM,kBAAA,GAAuC,OAAO;CACzE,aAAa,iBAAiB,KAAK;CACnC,MAAM,EAAE,OAAO,WAAW,KAAK,MAAM,aAAa,OAAO,aAAa;CACtE,IAAI,CAAC,IAAI,KAAK,KAAK,CAAC,IAAI,SAAS,KAAK,CAAC,IAAI,IAAI,KAAK,CAAC,IAAI,WAAW,KAAK,CAAC,IAAI,GAAG,GAAG,OAAO;CAC3F,IAAI,CAAC,QAAQ,QAAQ,KAAK,KAAK,CAAC,MAAM,QAAQ,QAAQ,GAAG,OAAO;CAChE,MAAM,OAAO,SAAS,IAAI,eAAe;CACzC,IAAI,CAAC,KAAK,OAAO,MAAM,MAAM,IAAI,GAAG,OAAO;CAC3C,OAAO;EACL,eAAA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,SAAS,KAAK;EACd,QAAQ,WAAW,IAAI;EACvB,UAAU;CACZ;AACF"}
|
|
@@ -1,5 +1,6 @@
|
|
|
1
1
|
import { n as UsageError } from "./errors-B6BSMs03.js";
|
|
2
|
-
import { a as
|
|
2
|
+
import { a as checkModelString, l as judgeRoute, o as isFakeJudge, s as judgeCredentials } from "./judge-Dt6KW6_O.js";
|
|
3
|
+
import { n as analysisModelOf } from "./config-Dv5MG7Rp.js";
|
|
3
4
|
import { z } from "zod";
|
|
4
5
|
import { Output, generateText } from "ai";
|
|
5
6
|
//#region src/analysis/schema.ts
|
|
@@ -85,12 +86,12 @@ Rules:
|
|
|
85
86
|
//#region src/analysis/model.ts
|
|
86
87
|
const isFake = (m) => typeof m === "object" && m !== null && m.kind === "fake-analysis";
|
|
87
88
|
const isEvaluationModel = (model) => model.startsWith("typesafe-ai/");
|
|
88
|
-
/** `analysis.model`, else the judge's when it's a language model
|
|
89
|
+
/** `reeval analyze`'s model: `analysis.model`, else the judge's when it's a language model. */
|
|
89
90
|
function defaultModel(config) {
|
|
90
|
-
const model = config
|
|
91
|
+
const model = analysisModelOf(config);
|
|
91
92
|
if (model !== void 0) return model;
|
|
92
93
|
const judge = config.judge.model;
|
|
93
|
-
|
|
94
|
+
return typeof judge === "string" && !isEvaluationModel(judge) ? judge : void 0;
|
|
94
95
|
}
|
|
95
96
|
/** ` (the judge, typesafe-ai/jev-1.14.0, is an evaluation model: it scores yes/no questions and can't <task>)`, when it is one. */
|
|
96
97
|
function judgeAside(config, task) {
|
|
@@ -99,10 +100,10 @@ function judgeAside(config, task) {
|
|
|
99
100
|
return ` (the judge, ${isFakeJudge(judge) ? "a fake judge" : judge}, is an evaluation model: it scores yes/no questions and can't ${task})`;
|
|
100
101
|
}
|
|
101
102
|
/** Which model analyses: `analysis.model`, else the judge's when it's a language model. */
|
|
102
|
-
function analysisModelName(config
|
|
103
|
+
function analysisModelName(config) {
|
|
103
104
|
const model = defaultModel(config);
|
|
104
105
|
if (model !== void 0) return model;
|
|
105
|
-
throw new UsageError(
|
|
106
|
+
throw new UsageError(`reeval analyze needs a language model. Set analysis: { model: 'google/gemini-3.8-flash' } in ${config.file}${judgeAside(config, "write an analysis")}, or run reeval analyze --context to hand the digest to your own coding agent.`);
|
|
106
107
|
}
|
|
107
108
|
/** The plan's name for the model a compare question without its own is answered by; null when none. */
|
|
108
109
|
function questionModelLabel(config) {
|
|
@@ -122,9 +123,19 @@ function questionModel(config, question, where) {
|
|
|
122
123
|
if (typeof model === "string" && isEvaluationModel(model)) throw new UsageError(`compare.questions "${question.question}" (${where}) would be answered by analysis.model in ${config.file}, "${model}", an evaluation model: it scores yes/no questions and can't pick between two attempts. Use a language model, e.g. 'google/gemini-3.8-flash', or ${fix}.`);
|
|
123
124
|
return model;
|
|
124
125
|
}
|
|
126
|
+
/**
|
|
127
|
+
* The variable `analysis.model` is missing when it's a provider model with no credential set: its
|
|
128
|
+
* provider's own key, else the AI Gateway's. Null when one is set, or for any other model.
|
|
129
|
+
*/
|
|
130
|
+
function missingAnalysisKey(config, env = process.env) {
|
|
131
|
+
const model = analysisModelOf(config);
|
|
132
|
+
if (typeof model !== "string" || isEvaluationModel(model) || checkModelString(model)) return null;
|
|
133
|
+
const keys = judgeCredentials(model);
|
|
134
|
+
return keys.some((k) => env[k]) ? null : keys[0] ?? null;
|
|
135
|
+
}
|
|
125
136
|
/** Checks the model and its credentials before any work; throws a UsageError (exit 2). */
|
|
126
|
-
function resolveAnalysisModel(config
|
|
127
|
-
const model = analysisModelName(config
|
|
137
|
+
function resolveAnalysisModel(config) {
|
|
138
|
+
const model = analysisModelName(config);
|
|
128
139
|
if (isFake(model)) return {
|
|
129
140
|
name: "fake",
|
|
130
141
|
route: "fake",
|
|
@@ -138,37 +149,52 @@ function resolveAnalysisModel(config, use = "analyze") {
|
|
|
138
149
|
};
|
|
139
150
|
} catch (error) {
|
|
140
151
|
const message = error.message.replace("judge model", "analysis model");
|
|
141
|
-
throw new UsageError(
|
|
152
|
+
throw new UsageError(`${message} Or run reeval analyze --context, which needs no key.`);
|
|
142
153
|
}
|
|
143
154
|
}
|
|
144
|
-
/**
|
|
145
|
-
|
|
155
|
+
/**
|
|
156
|
+
* A language model that answers `text`, through the AI SDK's own parsing, or throws `error`.
|
|
157
|
+
* With `countTokens`, it reports a token per 4 characters of its prompt and its answer, so a
|
|
158
|
+
* token budget applies to it as it would to a real model; otherwise it reports none.
|
|
159
|
+
*/
|
|
160
|
+
function cannedModel(answer, countTokens = false) {
|
|
146
161
|
return {
|
|
147
162
|
specificationVersion: "v4",
|
|
148
163
|
provider: "fake",
|
|
149
164
|
modelId: "fake-analysis",
|
|
150
165
|
supportedUrls: {},
|
|
151
|
-
doGenerate: async () => {
|
|
166
|
+
doGenerate: async (options) => {
|
|
152
167
|
if ("error" in answer) throw new Error(answer.error);
|
|
168
|
+
const input = countTokens ? Math.ceil(JSON.stringify(options.prompt).length / 4) : 0;
|
|
169
|
+
let output = countTokens ? Math.ceil(answer.text.length / 4) : 0;
|
|
170
|
+
let text = answer.text;
|
|
171
|
+
const cut = countTokens && options.maxOutputTokens !== void 0 && output > options.maxOutputTokens;
|
|
172
|
+
if (cut) {
|
|
173
|
+
output = options.maxOutputTokens;
|
|
174
|
+
text = text.slice(0, output * 4);
|
|
175
|
+
}
|
|
153
176
|
return {
|
|
154
177
|
content: [{
|
|
155
178
|
type: "text",
|
|
156
|
-
text
|
|
179
|
+
text
|
|
157
180
|
}],
|
|
158
|
-
finishReason: {
|
|
181
|
+
finishReason: cut ? {
|
|
182
|
+
unified: "length",
|
|
183
|
+
raw: "max_tokens"
|
|
184
|
+
} : {
|
|
159
185
|
unified: "stop",
|
|
160
186
|
raw: "stop"
|
|
161
187
|
},
|
|
162
188
|
usage: {
|
|
163
189
|
inputTokens: {
|
|
164
|
-
total:
|
|
165
|
-
noCache:
|
|
190
|
+
total: input,
|
|
191
|
+
noCache: input,
|
|
166
192
|
cacheRead: 0,
|
|
167
193
|
cacheWrite: 0
|
|
168
194
|
},
|
|
169
195
|
outputTokens: {
|
|
170
|
-
total:
|
|
171
|
-
text:
|
|
196
|
+
total: output,
|
|
197
|
+
text: output,
|
|
172
198
|
reasoning: 0
|
|
173
199
|
}
|
|
174
200
|
},
|
|
@@ -227,6 +253,6 @@ async function callAnalysisModel(model, digest, prompt, maxOutputTokens) {
|
|
|
227
253
|
};
|
|
228
254
|
}
|
|
229
255
|
//#endregion
|
|
230
|
-
export {
|
|
256
|
+
export { missingAnalysisKey as a, questionModelLabel as c, isFake as i, resolveAnalysisModel as l, callAnalysisModel as n, providerModel as o, cannedModel as r, questionModel as s, analysisModelName as t, SYSTEM_PROMPT as u };
|
|
231
257
|
|
|
232
|
-
//# sourceMappingURL=model-
|
|
258
|
+
//# sourceMappingURL=model-B-mel-Gv.js.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"model-B-mel-Gv.js","names":[],"sources":["../src/analysis/schema.ts","../src/analysis/model.ts"],"mappings":";;;;;;AAGA,MAAM,aAAa,EAAE,KAAK;CAAC;CAAQ;CAAU;AAAK,CAAC;AACnD,MAAM,QAAQ,EACX,MAAM,EAAE,OAAO,CAAC,CAAC,CACjB,SAAS,uFAAuF;AAEnG,MAAM,QAAQ,EAAE,OAAO;CACrB,MAAM,EAAE,OAAO;CACf;CACA;CACA,QAAQ,EACL,MAAM,EAAE,OAAO,CAAC,CAAC,CACjB,SAAS,CAAC,CACV,SAAS,qEAAqE;AACnF,CAAC;AAED,MAAa,iBAAiB,EAAE,OAAO;CACrC,MAAM,EACH,OAAO,CAAC,CACR,SACC,2HACF;CACF,cAAc,EAAE,MAAM,KAAK;CAC3B,SAAS,EACN,MAAM,KAAK,CAAC,CACZ,SACC,uLACF;CACF,KAAK,EACF,MAAM,KAAK,CAAC,CACZ,SAAS,yEAAyE;CACrF,WAAW,EAAE,MACX,EAAE,OAAO;EACP,MAAM,EAAE,OAAO;EACf,MAAM,EAAE,KAAK;GAAC;GAAc;GAAe;GAAS;GAAS;EAAO,CAAC;EACrE;CACF,CAAC,CACH;CACA,SAAS,EACN,MACC,EAAE,OAAO;EACP,OAAO,EAAE,OAAO;EAChB,QAAQ,EAAE,OAAO,CAAC,CAAC,SAAS,mDAAmD;EAC/E,MAAM,EAAE,KAAK;GAAC;GAAS;GAAQ;GAAQ;GAAS;GAAU;EAAS,CAAC;EACpE,QAAQ,EAAE,OAAO;EACjB,gBAAgB,EAAE,OAAO,CAAC,CAAC,SAAS,+BAA+B;EACnE,QAAQ,EACL,OAAO,CAAC,CACR,SACC,sIACF;EACF,QAAQ,EAAE,KAAK;GAAC;GAAQ;GAAU;EAAK,CAAC;EACxC,QAAQ,EAAE,KAAK;GAAC;GAAO;GAAU;EAAM,CAAC;EACxC;CACF,CAAC,CACH,CAAC,CACA,IAAI,CAAC;CACR,YAAY,EAAE,OAAO;EAAE,SAAS;EAAY,KAAK,EAAE,OAAO;CAAE,CAAC;AAC/D,CAAC;AAID,MAAa,gBAAgB;;;;;;;;;;;;;;;;;;;;;AChD7B,MAAa,UAAU,MACrB,OAAO,MAAM,YAAY,MAAM,QAAS,EAAwB,SAAS;AAE3E,MAAM,qBAAqB,UAAkB,MAAM,WAAW,cAAc;;AAG5E,SAAS,aAAa,QAAgE;CACpF,MAAM,QAAQ,gBAAgB,MAAM;CACpC,IAAI,UAAU,KAAA,GAAW,OAAO;CAChC,MAAM,QAAQ,OAAO,MAAM;CAC3B,OAAO,OAAO,UAAU,YAAY,CAAC,kBAAkB,KAAK,IAAI,QAAQ,KAAA;AAC1E;;AAGA,SAAS,WAAW,QAAwB,MAAsB;CAChE,MAAM,QAAQ,OAAO,MAAM;CAC3B,IAAI,OAAO,UAAU,YAAY,CAAC,YAAY,KAAK,GAAG,OAAO;CAC7D,OAAO,gBAAgB,YAAY,KAAK,IAAI,iBAAiB,MAAM,iEAAiE,KAAK;AAC3I;;AAGA,SAAgB,kBAAkB,QAAoD;CACpF,MAAM,QAAQ,aAAa,MAAM;CACjC,IAAI,UAAU,KAAA,GAAW,OAAO;CAChC,MAAM,IAAI,WACR,gGAAgG,OAAO,OAAO,WAAW,QAAQ,mBAAmB,EAAE,+EACxJ;AACF;;AAGA,SAAgB,mBAAmB,QAAuC;CACxE,MAAM,QAAQ,aAAa,MAAM;CACjC,IAAI,UAAU,KAAA,GAAW,OAAO;CAChC,OAAO,OAAO,KAAK,IAAI,SAAS;AAClC;;;;;AAMA,SAAgB,cACd,QACA,UACA,OAC4B;CAC5B,IAAI,SAAS,UAAU,KAAA,GAAW,OAAO,SAAS;CAClD,MAAM,QAAQ,aAAa,MAAM;CACjC,MAAM,MAAM;CACZ,IAAI,UAAU,KAAA,GACZ,MAAM,IAAI,WACR,sBAAsB,SAAS,SAAS,KAAK,MAAM,+FAA+F,OAAO,OAAO,WAAW,QAAQ,2BAA2B,EAAE,OAAO,IAAI,EAC7N;CAEF,IAAI,OAAO,UAAU,YAAY,kBAAkB,KAAK,GACtD,MAAM,IAAI,WACR,sBAAsB,SAAS,SAAS,KAAK,MAAM,2CAA2C,OAAO,KAAK,KAAK,MAAM,mJAAmJ,IAAI,EAC9Q;CAEF,OAAO;AACT;;;;;AAMA,SAAgB,mBACd,QACA,MAAyB,QAAQ,KAClB;CACf,MAAM,QAAQ,gBAAgB,MAAM;CACpC,IAAI,OAAO,UAAU,YAAY,kBAAkB,KAAK,KAAK,iBAAiB,KAAK,GAAG,OAAO;CAC7F,MAAM,OAAO,iBAAiB,KAAK;CACnC,OAAO,KAAK,MAAM,MAAM,IAAI,EAAE,IAAI,OAAQ,KAAK,MAAM;AACvD;;AAGA,SAAgB,qBAAqB,QAA+C;CAClF,MAAM,QAAQ,kBAAkB,MAAM;CACtC,IAAI,OAAO,KAAK,GAAG,OAAO;EAAE,MAAM;EAAQ,OAAO;EAAQ,MAAM;CAAM;CACrE,IAAI,kBAAkB,KAAK,GACzB,MAAM,IAAI,WACR,qBAAqB,OAAO,KAAK,OAAO,MAAM,sIAChD;CAEF,IAAI;EACF,OAAO;GAAE,MAAM;GAAO,OAAO,WAAW,KAAK,CAAC,CAAC;EAAK;CACtD,SAAS,OAAO;EACd,MAAM,UAAW,MAAgB,QAAQ,QAAQ,eAAe,gBAAgB;EAChF,MAAM,IAAI,WAAW,GAAG,QAAQ,sDAAsD;CACxF;AACF;;;;;;AAOA,SAAgB,YACd,QACA,cAAc,OACC;CACf,OAAO;EACL,sBAAsB;EACtB,UAAU;EACV,SAAS;EACT,eAAe,CAAC;EAChB,YAAY,OAAO,YAAY;GAC7B,IAAI,WAAW,QAAQ,MAAM,IAAI,MAAM,OAAO,KAAK;GACnD,MAAM,QAAQ,cAAc,KAAK,KAAK,KAAK,UAAU,QAAQ,MAAM,CAAC,CAAC,SAAS,CAAC,IAAI;GACnF,IAAI,SAAS,cAAc,KAAK,KAAK,OAAO,KAAK,SAAS,CAAC,IAAI;GAC/D,IAAI,OAAO,OAAO;GAElB,MAAM,MACJ,eAAe,QAAQ,oBAAoB,KAAA,KAAa,SAAS,QAAQ;GAC3E,IAAI,KAAK;IACP,SAAS,QAAQ;IACjB,OAAO,KAAK,MAAM,GAAG,SAAS,CAAC;GACjC;GACA,OAAO;IACL,SAAS,CAAC;KAAE,MAAM;KAAQ;IAAK,CAAC;IAChC,cAAc,MACV;KAAE,SAAS;KAAU,KAAK;IAAa,IACvC;KAAE,SAAS;KAAQ,KAAK;IAAO;IACnC,OAAO;KACL,aAAa;MAAE,OAAO;MAAO,SAAS;MAAO,WAAW;MAAG,YAAY;KAAE;KACzE,cAAc;MAAE,OAAO;MAAQ,MAAM;MAAQ,WAAW;KAAE;IAC5D;IACA,UAAU,CAAC;IACX,UAAU,EAAE,SAAS,gBAAgB;GACvC;EACF;EACA,UAAU,YAAY;GACpB,MAAM,IAAI,MAAM,yCAAyC;EAC3D;CACF;AACF;;AAGA,SAAS,UAAU,MAAyB,QAA+B;CACzE,IAAI,KAAK,WAAW,KAAA,GAClB,MAAM,IAAI,MAAM,+DAA+D;CAEjF,MAAM,SAAS,OAAO,KAAK,WAAW,aAAa,KAAK,OAAO,MAAM,IAAI,KAAK;CAC9E,OAAO,YAAY,EAAE,MAAM,OAAO,WAAW,WAAW,SAAS,KAAK,UAAU,MAAM,EAAE,CAAC;AAC3F;AAEA,eAAe,cAAc,OAA8B,QAAwC;CACjG,IAAI,MAAM,MAAM,OAAO,UAAU,MAAM,MAAM,MAAM;CACnD,OAAO,cAAc,KAAK;AAC5B;;AAGA,eAAsB,cAAc,OAAsD;CACxF,IAAI,MAAM,UAAU,WAAW,OAAO,MAAM;CAC5C,MAAM,WAAW,MAAM,KAAK,MAAM,GAAG,MAAM,KAAK,QAAQ,GAAG,CAAC;CAC5D,MAAM,KAAK,MAAM,KAAK,MAAM,SAAS,SAAS,CAAC;CAC/C,IAAI,aAAa,UAAU;EACzB,MAAM,EAAE,WAAW,MAAM,OAAO;EAChC,OAAO,OAAO,EAAE;CAClB;CACA,OAAO,MAAM;AACf;AASA,eAAsB,kBACpB,OACA,QACA,QACA,iBACsB;CACtB,MAAM,UAAU,KAAK,IAAI;CACzB,MAAM,SAAS,MAAM,aAAa;EAChC,OAAO,MAAM,cAAc,OAAO,MAAM;EACxC,QAAQ;EACR;EACA,QAAQ,OAAO,OAAO;GAAE,QAAQ;GAAgB,MAAM;EAAW,CAAC;EAClE,aAAa;EACb;EACA,YAAY;CACd,CAAC;CACD,OAAO;EACL,QAAQ,OAAO;EACf,eAAe,OAAO,UAAU,WAAW;EAC3C,OAAO;GACL,aAAa,OAAO,OAAO,eAAe;GAC1C,cAAc,OAAO,OAAO,gBAAgB;EAC9C;EACA,WAAW,KAAK,IAAI,IAAI;CAC1B;AACF"}
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
export { t as analysisModelName, n as callAnalysisModel, r as cannedModel, i as isFake, a as missingAnalysisKey, o as providerModel, s as questionModel, c as questionModelLabel, l as resolveAnalysisModel } from "./model-B-mel-Gv.js";
|
|
@@ -1,4 +1,4 @@
|
|
|
1
|
-
import {
|
|
1
|
+
import { l as judgeRoute } from "./judge-Dt6KW6_O.js";
|
|
2
2
|
import { i as debugTrace } from "./attempt-CGPV3BZY.js";
|
|
3
3
|
import { experimental_decide } from "ai";
|
|
4
4
|
import { google } from "@ai-sdk/google";
|
|
@@ -41,4 +41,4 @@ async function scoreWithModel(model, question, state, providerOptions, signal) {
|
|
|
41
41
|
//#endregion
|
|
42
42
|
export { scoreWithModel };
|
|
43
43
|
|
|
44
|
-
//# sourceMappingURL=model-
|
|
44
|
+
//# sourceMappingURL=model-C7tSwT7f.js.map
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"model-
|
|
1
|
+
{"version":3,"file":"model-C7tSwT7f.js","names":["decide"],"sources":["../src/judge/model.ts"],"mappings":";;;;;;AASA,eAAe,aACb,OACwB;CACxB,IAAI,OAAO,UAAU,UAAU,OAAO;CACtC,MAAM,QAAQ,WAAW,KAAK;CAC9B,IAAI,MAAM,SAAS,WAAW,OAAO;CACrC,MAAM,KAAK,MAAM,MAAM,MAAM,SAAS,SAAS,CAAC;CAChD,IAAI,MAAM,aAAa,UAAU,OAAO,OAAO,cAAc,EAAE;CAC/D,MAAM,EAAE,eAAe,MAAM,OAAO;CACpC,OAAO,WAAW,cAAc,EAAE;AACpC;AAEA,eAAsB,eACpB,OACA,UACA,OACA,iBACA,QAC+D;CAC/D,MAAM,SAAS,MAAMA,oBAAO;EAC1B,OAAO,MAAM,aAAa,KAAK;EAC/B;EACA,WAAW,EAAE,QAAQ;GAAE,MAAM;GAAW,cAAc;EAAS,EAAE;EACjE;EACA,aAAa;EAEb,YAAY;CACd,CAAC;CACD,MAAM,EAAE,SAAS,OAAO,UAAU,UAAU,qBAAqB;CACjE,WACE,SACA,qBACA;EAAE;EAAS;EAAO;EAAU;EAAU;CAAiB,GACvD,yBACF;CACA,OAAO;EAAE,OAAO,OAAO,QAAQ,OAAO;EAAa,eAAe,OAAO,UAAU;CAAQ;AAC7F"}
|