Files
confidence-engine/scripts/experimental/rto-focused-vs-global-comparison.mjs
T

309 lines
11 KiB
JavaScript

import fs from "fs/promises";
import path from "path";
import dotenv from "dotenv";
import fixture from "../../tests/fixtures/live-product-launch-update-response.json" with { type: "json" };
import { buildGraphUpdatePrompt } from "../../lib/graph/prompt-builder.js";
import { assertConfig } from "../../lib/config.js";
import { createRequire } from "module";
const require = createRequire(import.meta.url);
const { runLiveExperiment } = require("../../tests/graph/live-update-experiment-helper.cjs");
// Non-production diagnostics: an independent request captures raw-response
// text and Ollama completion metadata (done, done_reason, prompt_eval_count,
// eval_count) before delegating JSON recovery to the inline helper.
import { runFocusedDiagnostic } from "./focus-diagnostics-helper.mjs";
dotenv.config({ path: ".env.local" });
const TARGET_NODE_ID = "nxmeiab";
const FIXED_QUESTION =
"What evidence would clarify whether competitors are actively developing similar products and how soon they might release them?";
const FIXED_ANSWER =
"Two competitors have publicly announced products aimed at the same customer problem. One says it expects a beta within six months, while the other has not announced a release date. We do not yet know how closely either product matches ours.";
function getGraph() {
return structuredClone(fixture.updatedSituationGraph);
}
function getTargetNode(graph) {
const targetNode = graph.nodes.find((node) => node.id === TARGET_NODE_ID);
if (!targetNode || targetNode.kind !== "unknown" || targetNode.status === "resolved") {
throw new Error("Fixed target node nxmeiab is not present as unresolved unknown in fixture.");
}
return targetNode;
}
function buildFocusedPrompt({ targetNode, centralStatement, question, answer }) {
return `You are performing focused answer deconstruction for one explicitly user-chosen investigation.
Return exactly one JSON object. Return JSON only.
This is NOT a graph update task.
Do NOT output graph mutations.
Do NOT output selection, ranking, ownership, recommendation, confidence, or next-best-question semantics.
Do NOT include any of these fields: addedNodes, updatedNodes, removedNodes, addedEdges, removedEdges, resolvedNodeIds, activeUnknownNodeId, selectedQuestion.
Required top-level fields:
- targetNodeId
- observations
- uncertainties
- assumptions
- relationships
- possibleFollowUpQuestions
Field rules:
- targetNodeId must be exactly "${TARGET_NODE_ID}"
- observations: only statements directly supported by the answer
- uncertainties: only things the answer explicitly leaves unknown or unclear
- assumptions: include only if the answer itself relies on an assumption
- relationships: only direct supported relationships among extracted items, each with { from, to, type, rationale }
- possibleFollowUpQuestions: unresolved questions genuinely exposed by this answer, unranked
Focused case context:
- target label: ${targetNode.label}
- target description: ${targetNode.description}
- central case statement: ${centralStatement}
Question:
${question}
Answer:
${answer}`;
}
function assertRequiredEnv(name) {
const value = process.env[name];
if (!value) {
throw new Error(`${name} is required in .env.local`);
}
return value;
}
function validateEnvironment() {
const baseUrl = assertRequiredEnv("OLLAMA_BASE_URL");
if (baseUrl === "http://localhost:11434" || baseUrl === "http://127.0.0.1:11434") {
throw new Error(`Comparison apparatus refuses localhost fallback. OLLAMA_BASE_URL=${baseUrl}`);
}
const config = assertConfig();
return { baseUrl, modelName: config.OLLAMA_MODEL };
}
function buildFocusedPlan() {
const graph = getGraph();
const targetNode = getTargetNode(graph);
const prompt = buildFocusedPrompt({
targetNode,
centralStatement: graph.centralStatement,
question: FIXED_QUESTION,
answer: FIXED_ANSWER,
});
return {
path: "focused",
targetNodeId: TARGET_NODE_ID,
modelTimingBoundary:
"Immediately before provider.generateReconstruction(prompt, modelName) to immediately after it resolves.",
inputCharacterCount: prompt.length,
prompt,
focusedContextSummary: [
"target node id and label",
"target node description",
"central case statement",
"fixed question",
"fixed answer",
],
};
}
function buildGlobalPlan() {
const graph = getGraph();
const prompt = buildGraphUpdatePrompt({
situationGraph: graph,
previousQuestion: FIXED_QUESTION,
answer: FIXED_ANSWER,
});
return {
path: "global",
targetNodeId: TARGET_NODE_ID,
modelTimingBoundary:
"Immediately before updateCase() calls provider.generateReconstruction(prompt, modelName) to immediately after it resolves inside orchestrator.",
endToEndTimingBoundary:
"Immediately before updateCase() entry to immediately after full updateCase() result returns.",
inputCharacterCount: prompt.length,
prompt,
graph,
};
}
async function runFocused() {
const { baseUrl, modelName } = validateEnvironment();
const plan = buildFocusedPlan();
// ---- non-production diagnostics wrapper ----
// Independent request captures raw response + completion metadata.
// The diagnostics helper mirrors the production JSON-recovery logic inline
// (since recoverJson is not exported) and delegates parse back to that
// implementation. The comparison runner only observes result/error objects.
const diagResult = await runFocusedDiagnostic({ baseUrl, modelName, prompt: plan.prompt });
try {
if (diagResult.error) {
throw diagResult.error;
}
// Successful parse via production helper — proceed to standard artifact
const raw = diagResult.parsedResult;
const artifact = {
path: "focused",
targetNodeId: TARGET_NODE_ID,
modelName,
modelElapsedMs: diagResult.diagnostics.modelElapsedMs,
inputCharacterCount: plan.inputCharacterCount,
structuredResult: raw,
};
const artifactPath = path.resolve(
"tests/experimental/artifacts/rto-focused-vs-global-focused.json",
);
await fs.mkdir(path.dirname(artifactPath), { recursive: true });
await fs.writeFile(artifactPath, JSON.stringify(artifact, null, 2));
console.log(JSON.stringify({ artifactPath, modelElapsedMs: diagResult.diagnostics.modelElapsedMs, inputCharacterCount: plan.inputCharacterCount }, null, 2));
return;
} catch (err) {
const focusedDiagnostics = err.focusedDiagnostics ?? null;
if (!focusedDiagnostics) {
// Should not happen — every DiagnosticError carries diagnostics.
console.error("Unexpected: no diagnostics captured before parse failure.");
process.exit(1);
}
const errorArtifact = {
path: "focused",
modelName: focusedDiagnostics.apiUsed === "/api/chat" ? modelName : modelName,
inputCharacterCount: plan.inputCharacterCount,
rawResponseText: focusedDiagnostics.rawResponseText?.slice(0, 4096),
rawResponseCharacterCount: focusedDiagnostics.rawResponseCharacterCount,
done: focusedDiagnostics.done,
doneReason: focusedDiagnostics.doneReason,
promptEvalCount: focusedDiagnostics.promptEvalCount,
evalCount: focusedDiagnostics.evalCount,
errorType: err.name,
errorMessage: err.message.slice(0, 4096),
};
const artifactPath = path.resolve(
"tests/experimental/artifacts/rto-focused-vs-global-focused-failure.json",
);
await fs.mkdir(path.dirname(artifactPath), { recursive: true });
await fs.writeFile(artifactPath, JSON.stringify(errorArtifact, null, 2));
console.log(JSON.stringify({
artifactPath,
errorType: err.name,
rawResponseCharacterCount: focusedDiagnostics.rawResponseCharacterCount,
done: focusedDiagnostics.done,
doneReason: focusedDiagnostics.doneReason,
}, null, 2));
}
}
async function runGlobal() {
const { modelName } = validateEnvironment();
const plan = buildGlobalPlan();
const startedAt = Date.now();
const result = await runLiveExperiment({
graph: plan.graph,
previousQuestion: FIXED_QUESTION,
answer: FIXED_ANSWER,
});
const endToEndElapsedMs = Date.now() - startedAt;
const artifact = {
path: "global",
targetNodeId: TARGET_NODE_ID,
modelName,
modelElapsedMs: result.modelElapsedMs,
providerGenerateCalls: result.providerGenerateCalls ?? 0,
providerTimingCaptured: typeof result.modelElapsedMs === "number",
endToEndElapsedMs: result.endToEndElapsedMs ?? endToEndElapsedMs,
inputCharacterCount: plan.inputCharacterCount,
graphUpdateResultSummary: {
userSupportedMeaning: result.userSupportedMeaning,
possibleInference: result.possibleInference,
proposalValidation: result.proposalValidation,
selectedQuestion: result.selectedQuestion,
},
answerDerivedComparisonExtraction: {
observations: result.userSupportedMeaning ? [result.userSupportedMeaning] : "not directly exposed",
uncertainties: result.possibleInference ? [result.possibleInference] : "not directly exposed",
assumptions: "not directly exposed",
relationships: "not directly exposed",
newlySurfacedUnknownsOrQuestions: result.selectedQuestion ?? "not directly exposed",
},
};
const artifactPath = path.resolve(
"tests/experimental/artifacts/rto-focused-vs-global-global.json",
);
await fs.mkdir(path.dirname(artifactPath), { recursive: true });
await fs.writeFile(artifactPath, JSON.stringify(artifact, null, 2));
console.log(JSON.stringify({ artifactPath, modelElapsedMs: result.modelElapsedMs, endToEndElapsedMs: result.endToEndElapsedMs ?? endToEndElapsedMs, inputCharacterCount: plan.inputCharacterCount }, null, 2));
}
function printStaticValidation() {
const { modelName } = validateEnvironment();
const focused = buildFocusedPlan();
const global = buildGlobalPlan();
const summary = {
modelName,
fixedTargetNodeId: TARGET_NODE_ID,
fixedQuestion: FIXED_QUESTION,
fixedAnswer: FIXED_ANSWER,
focused: {
resolvesTo: "scripts/experimental/rto-focused-vs-global-comparison-equivalent focused provider path",
wholeGraphSupplied: false,
inputCharacterCount: focused.inputCharacterCount,
modelTimingBoundary: focused.modelTimingBoundary,
},
global: {
resolvesTo: "tests/graph/live-update-experiment-helper.cjs -> updateCase() production path",
wholeGraphSupplied: true,
inputCharacterCount: global.inputCharacterCount,
modelTimingBoundary: global.modelTimingBoundary,
endToEndTimingBoundary: global.endToEndTimingBoundary,
providerTimingSource: "non-production timed provider wrapper injected through updateCase dependencies.provider",
},
commands: {
focused: "node scripts/experimental/rto-focused-vs-global-comparison.mjs --focused",
global: "node scripts/experimental/rto-focused-vs-global-comparison.mjs --global",
},
};
console.log(JSON.stringify(summary, null, 2));
}
const mode = process.argv[2] ?? "--validate";
if (mode === "--focused") {
runFocused().catch((error) => {
console.error(error instanceof Error ? error.message : String(error));
process.exit(1);
});
} else if (mode === "--global") {
runGlobal().catch((error) => {
console.error(error instanceof Error ? error.message : String(error));
process.exit(1);
});
} else if (mode === "--validate") {
printStaticValidation();
} else {
console.error(`Unknown mode: ${mode}`);
process.exit(1);
}