mirror of
https://github.com/OpenSquawk/OpenSquawk
synced 2026-08-01 22:26:04 +08:00
feat: integrate llm-backed routing with fallback
This commit is contained in:
Remi
@@ -874,144 +874,167 @@ function optimizeInputForLLM(input: LLMDecisionInput) {
|
||||
}
|
||||
|
||||
|
||||
function summarizeCandidateForPrompt(candidate: DecisionCandidate) {
|
||||
const { state } = candidate
|
||||
return {
|
||||
id: candidate.id,
|
||||
flow: candidate.flow,
|
||||
role: state?.role,
|
||||
phase: state?.phase,
|
||||
summary: state?.summary,
|
||||
say_tpl: state?.say_tpl,
|
||||
utterance_tpl: state?.utterance_tpl,
|
||||
handoff: state?.handoff,
|
||||
}
|
||||
}
|
||||
|
||||
function extractJsonObject(text: string): any | null {
|
||||
if (!text) return null
|
||||
const trimmed = text.trim()
|
||||
try {
|
||||
return JSON.parse(trimmed)
|
||||
} catch {}
|
||||
|
||||
const match = trimmed.match(/\{[\s\S]*\}/)
|
||||
if (!match) {
|
||||
return null
|
||||
}
|
||||
try {
|
||||
return JSON.parse(match[0])
|
||||
} catch {
|
||||
return null
|
||||
}
|
||||
}
|
||||
|
||||
export async function routeDecision(input: LLMDecisionInput): Promise<LLMDecisionResult> {
|
||||
const utterance = (input.pilot_utterance || '').trim()
|
||||
const { system, index } = await getRuntimeSystemIndex()
|
||||
const prepared = await prepareDecisionCandidates(input, utterance)
|
||||
|
||||
const currentEntry = index.get(input.state_id)
|
||||
const activeFlowSlug =
|
||||
input.flow_slug
|
||||
|| currentEntry?.flow
|
||||
|| system.main
|
||||
|| Object.keys(system.flows || {})[0]
|
||||
|| ''
|
||||
|
||||
const candidateMap = new Map<string, DecisionCandidate>()
|
||||
const addCandidate = (
|
||||
id?: string,
|
||||
flow?: string,
|
||||
providedState?: RuntimeDecisionState
|
||||
) => {
|
||||
if (!id || candidateMap.has(id)) {
|
||||
return
|
||||
}
|
||||
|
||||
const indexed = index.get(id)
|
||||
const state = providedState || indexed?.state
|
||||
if (!state) {
|
||||
return
|
||||
}
|
||||
|
||||
const triggers = Array.isArray(state.triggers) ? state.triggers.filter(Boolean) : []
|
||||
const regexTriggers = triggers.filter(trigger => trigger?.type === 'regex')
|
||||
const noneTriggers = triggers.filter(trigger => trigger?.type === 'none')
|
||||
const flowSlug = flow || indexed?.flow || activeFlowSlug
|
||||
|
||||
candidateMap.set(id, {
|
||||
id,
|
||||
flow: flowSlug,
|
||||
state,
|
||||
triggers,
|
||||
regexTriggers,
|
||||
noneTriggers,
|
||||
})
|
||||
const trace: LLMDecisionTrace = {
|
||||
calls: [],
|
||||
candidateTimeline: prepared.timeline,
|
||||
}
|
||||
|
||||
if (currentEntry?.state) {
|
||||
const transitions = [
|
||||
...(currentEntry.state.next || []),
|
||||
...(currentEntry.state.ok_next || []),
|
||||
...(currentEntry.state.bad_next || []),
|
||||
...(currentEntry.state.timer_next || []),
|
||||
]
|
||||
for (const transition of transitions) {
|
||||
if (!transition?.to) continue
|
||||
addCandidate(transition.to, currentEntry.flow)
|
||||
if (prepared.autoSelected) {
|
||||
trace.autoSelection = {
|
||||
id: prepared.autoSelected.id,
|
||||
flow: prepared.autoSelected.flow,
|
||||
reason: 'Heuristic routing resolved a single remaining candidate.',
|
||||
}
|
||||
return {
|
||||
decision: { next_state: prepared.autoSelected.id },
|
||||
trace,
|
||||
}
|
||||
}
|
||||
|
||||
for (const raw of input.candidates || []) {
|
||||
if (!raw?.id) continue
|
||||
addCandidate(raw.id, raw.flow, raw.state)
|
||||
}
|
||||
const candidatePool = prepared.finalCandidates.length > 0
|
||||
? prepared.finalCandidates
|
||||
: Array.from(prepared.candidateIndex.values())
|
||||
|
||||
for (const [flowSlug, tree] of Object.entries(system.flows || {})) {
|
||||
if (flowSlug === activeFlowSlug) continue
|
||||
const start = tree?.start_state
|
||||
if (!start) continue
|
||||
const startState = tree?.states?.[start]
|
||||
addCandidate(start, flowSlug, startState)
|
||||
}
|
||||
|
||||
let candidates = Array.from(candidateMap.values())
|
||||
if (candidates.length === 0) {
|
||||
return { decision: { next_state: input.state_id } }
|
||||
}
|
||||
|
||||
candidates = candidates.filter(candidate =>
|
||||
!candidate.triggers.some(trigger => trigger?.type === 'auto_time' || trigger?.type === 'auto_variable')
|
||||
)
|
||||
|
||||
const regexCandidates = candidates.filter(candidate => candidate.regexTriggers.length > 0)
|
||||
const regexMatches = regexCandidates.filter(candidate =>
|
||||
candidate.regexTriggers.some(trigger => evaluateRegexPattern(trigger.pattern, trigger.patternFlags, utterance))
|
||||
)
|
||||
|
||||
let trace: LLMDecisionTrace | undefined
|
||||
if (regexMatches.length > 0) {
|
||||
trace = { calls: [] }
|
||||
}
|
||||
|
||||
let workingSet = regexMatches
|
||||
if (workingSet.length === 0) {
|
||||
workingSet = candidates.filter(candidate => candidate.regexTriggers.length === 0)
|
||||
}
|
||||
|
||||
if (workingSet.length === 0) {
|
||||
return { decision: { next_state: input.state_id } }
|
||||
}
|
||||
|
||||
const context = { variables: input.variables || {}, flags: input.flags || {} }
|
||||
const survivors = workingSet.filter(candidate =>
|
||||
evaluateConditionList(candidate.state?.conditions, context, utterance).passed
|
||||
)
|
||||
|
||||
if (survivors.length === 1) {
|
||||
const [winner] = survivors
|
||||
|
||||
if (regexMatches.length > 0 && winner.regexTriggers.length > 0) {
|
||||
if (!trace) {
|
||||
trace = { calls: [] }
|
||||
}
|
||||
const patterns = winner.regexTriggers
|
||||
.map(trigger => trigger?.pattern ? `/${trigger.pattern}/${trigger.patternFlags || 'i'}` : '')
|
||||
.filter(pattern => Boolean(pattern))
|
||||
trace.autoSelection = {
|
||||
id: winner.id,
|
||||
flow: winner.flow,
|
||||
reason: patterns.length
|
||||
? `Regex trigger matched ${patterns.join(', ')}`
|
||||
: 'Regex trigger matched pilot utterance'
|
||||
}
|
||||
return { decision: { next_state: winner.id }, trace }
|
||||
}
|
||||
|
||||
return { decision: { next_state: winner.id } }
|
||||
}
|
||||
|
||||
if (survivors.length === 0) {
|
||||
return { decision: { next_state: input.state_id } }
|
||||
}
|
||||
|
||||
const [first] = survivors
|
||||
if (trace) {
|
||||
if (candidatePool.length === 0) {
|
||||
const fallbackState = fallbackNextState(input)
|
||||
trace.fallback = {
|
||||
used: true,
|
||||
reason: 'Multiple candidates matched after filtering; defaulting to first match.',
|
||||
selected: first.id,
|
||||
reason: 'No viable candidates after heuristic evaluation; falling back to default transition.',
|
||||
selected: fallbackState,
|
||||
}
|
||||
return { decision: { next_state: first.id }, trace }
|
||||
return { decision: { next_state: fallbackState }, trace }
|
||||
}
|
||||
|
||||
return { decision: { next_state: first.id } }
|
||||
const optimizedInput = optimizeInputForLLM({
|
||||
...input,
|
||||
candidates: candidatePool.map(candidate => ({
|
||||
id: candidate.id,
|
||||
flow: candidate.flow,
|
||||
state: candidate.state,
|
||||
})),
|
||||
})
|
||||
|
||||
const candidateSummaries = candidatePool
|
||||
.map(candidate => {
|
||||
const summary = [
|
||||
`${candidate.id}`,
|
||||
candidate.state?.summary || candidate.state?.say_tpl || candidate.state?.utterance_tpl || '',
|
||||
]
|
||||
.filter(Boolean)
|
||||
.join(' — ')
|
||||
return `- ${summary}`
|
||||
})
|
||||
.join('\n')
|
||||
|
||||
const systemPrompt = [
|
||||
'You are an assistant that selects the correct next state in an aviation decision tree.',
|
||||
'Evaluate the pilot transmission and choose the most appropriate candidate state id from the provided list.',
|
||||
'Respond strictly with a JSON object: {"next_state": "STATE_ID", "reason": "short rationale"}.',
|
||||
'Only use state ids that were provided. If you cannot decide, choose the best heuristic option.',
|
||||
].join(' ')
|
||||
|
||||
const userPrompt = [
|
||||
`Pilot transmission: "${utterance || '(silence)'}"`,
|
||||
'Candidate options:',
|
||||
candidateSummaries,
|
||||
'Context (JSON):',
|
||||
JSON.stringify(optimizedInput, null, 2),
|
||||
].join('\n')
|
||||
|
||||
const callEntry = {
|
||||
stage: 'decision' as const,
|
||||
request: {
|
||||
systemPrompt,
|
||||
userPrompt,
|
||||
candidates: candidatePool.map(summarizeCandidateForPrompt),
|
||||
},
|
||||
}
|
||||
trace.calls.push(callEntry)
|
||||
|
||||
try {
|
||||
const rawResponse = await decide(systemPrompt, userPrompt)
|
||||
callEntry.rawResponseText = rawResponse
|
||||
const parsed = extractJsonObject(rawResponse)
|
||||
if (parsed && typeof parsed.next_state === 'string' && parsed.next_state.trim().length > 0) {
|
||||
callEntry.response = parsed
|
||||
|
||||
const resolved =
|
||||
prepared.finalCandidateIndex.get(parsed.next_state)
|
||||
|| prepared.candidateIndex.get(parsed.next_state)
|
||||
|
||||
if (resolved) {
|
||||
return {
|
||||
decision: { next_state: resolved.id },
|
||||
trace,
|
||||
}
|
||||
}
|
||||
|
||||
return {
|
||||
decision: { next_state: parsed.next_state },
|
||||
trace,
|
||||
}
|
||||
}
|
||||
|
||||
throw new Error('LLM response missing next_state field')
|
||||
} catch (err: any) {
|
||||
callEntry.error = err?.message || String(err)
|
||||
|
||||
const fallbackCandidate = prepared.finalCandidates[0] || candidatePool[0]
|
||||
const fallbackState = fallbackCandidate?.id || fallbackNextState(input)
|
||||
|
||||
trace.fallback = {
|
||||
used: true,
|
||||
reason: 'OpenAI decision failed or was inconclusive; falling back to heuristic selection.',
|
||||
selected: fallbackState,
|
||||
}
|
||||
|
||||
return {
|
||||
decision: { next_state: fallbackState },
|
||||
trace,
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
export function __setRuntimeDecisionSystemForTests(system: RuntimeDecisionSystem) {
|
||||
runtimeSystemCache = {
|
||||
system,
|
||||
index: buildRuntimeIndex(system),
|
||||
timestamp: Date.now(),
|
||||
}
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user