// @harn-entrypoint-category llm.stdlib
//
// std/llm/ensemble — multi-sample helpers (best_of_n, self_consistency,
// parallel_judge, debate). Each function uses safe_call from std/llm/safe so
// failures surface as {ok: false, status} envelopes rather than throws.
//
// Citations:
// - Wang et al. 2022 "Self-Consistency Improves Chain of Thought Reasoning"
// (arxiv:2203.11171).
// - Du et al. 2023 "Improving Factuality and Reasoning in Language Models
// through Multiagent Debate" (arxiv:2305.14325).
// - OpenAI best-of-n / reranking patterns (cookbook).
import { agent_emit_event } from "std/agent/state"
import { safe_call, with_case_insensitive_keys } from "std/llm/safe"
fn __clamp_int(n, lo, hi) {
const v = to_int(n)
if v < lo {
return lo
}
if v > hi {
return hi
}
return v
}
fn __sampler_opts_for(opts) {
const so = opts?.sampler_opts
if type_of(so) == "dict" {
return so
}
return {temperature: 1.0}
}
fn __merge_call_opts(base, sampler_opts) {
const merged = if type_of(base) == "dict" {
base
} else {
{}
}
return merged + sampler_opts
}
// -------------------------------------------------------------------------------------------------
// best_of_n
// -------------------------------------------------------------------------------------------------
fn __sample_one(prompt, system, call_opts, index) {
const envelope = safe_call(prompt, system, call_opts)
if !(envelope?.ok ?? false) {
return {ok: false, index: index, status: envelope?.status, error: envelope?.error, text: ""}
}
const text = to_string(envelope.value?.text ?? "")
return {ok: true, index: index, text: text, value: envelope.value}
}
fn __collect_samples(prompt, system, base_opts, sampler_opts, n, run_parallel) {
const call_opts = __merge_call_opts(base_opts, sampler_opts)
let indices = []
let i = 0
while i < n {
indices = indices.push(i)
i = i + 1
}
if run_parallel {
return parallel each indices { idx ->
__sample_one(prompt, system, call_opts, idx)
}
}
let out = []
for idx in indices {
out = out.push(__sample_one(prompt, system, call_opts, idx))
}
return out
}
fn __reward_filter(samples, reward) {
if reward == nil {
return samples
}
let scored = []
for s in samples {
if s?.ok ?? false {
const r = try {
reward(s.text)
}
if !is_err(r) {
scored = scored.push(s + {reward: to_float(unwrap(r))})
} else {
scored = scored.push(s + {reward: 0.0})
}
}
}
return scored
}
fn __structured_judge(samples, opts) {
let labelled = []
const total = len(samples)
let i = 0
while i < total {
const s = samples[i]
if s?.ok ?? false {
labelled = labelled.push("[" + to_string(i) + "] " + to_string(s.text))
}
i = i + 1
}
if len(labelled) == 0 {
return {ok: false, status: "no_valid_samples"}
}
const prompt = if opts?.judge_prompt != nil && to_string(opts.judge_prompt) != "" {
to_string(opts.judge_prompt) + "\n\n" + join(labelled, "\n\n")
} else {
"Given these candidate answers, pick the best one. Reply ONLY with"
+ " a JSON object {\"best_index\": <int>, \"scores\": [..],"
+ " \"reasoning\": \"<brief>\"}.\n\n"
+ join(labelled, "\n\n")
}
const system = opts?.judge_system ?? ""
let jopts = if type_of(opts?.judge_opts) == "dict" {
opts.judge_opts
} else {
{}
}
if opts?.judge_model != nil && to_string(opts.judge_model) != "" {
jopts = jopts + {model: to_string(opts.judge_model)}
}
const schema = {
type: "object",
properties: {
best_index: {type: "integer"},
scores: {type: "array", items: {type: "number"}},
reasoning: {type: "string"},
},
required: ["best_index"],
}
const envelope = try {
llm_call_structured_result(prompt, schema, jopts + {system: system})
}
if is_err(envelope) {
return {ok: false, status: "exception", error: unwrap_err(envelope)}
}
const env = unwrap(envelope)
if !(env?.ok ?? false) {
return {ok: false, status: env?.status ?? "judge_failed", error: env?.error}
}
const data = with_case_insensitive_keys(env?.data ?? {})
const best = to_int(data?.best_index ?? 0)
return {
ok: true,
best_index: best,
scores: data?.scores ?? [],
reasoning: to_string(data?.reasoning ?? ""),
}
}
/**
* best_of_n(prompt, system, opts) -> dict
*
* Sample N candidates in parallel and ask a judge to pick the best.
*
* Options:
* - n: int (default 5; clamped to [2, 32])
* - sampler_opts: dict (default {temperature: 1.0})
* - judge: "structured" | closure (default "structured")
* - judge_opts, judge_model, judge_prompt, judge_system: optional
* - parallel: bool (default true)
* - reward: closure(text) -> float (optional pre-judge filter)
*
* Returns: {ok, best: {text, index, score?}, candidates, judge, reasoning}.
* On all-fail: {ok: false, status: "all_samples_failed"}.
*
* @effects: []
* @errors: []
*/
pub fn best_of_n(prompt, system, opts = nil) {
const cfg = if type_of(opts) == "dict" {
opts
} else {
{}
}
const n = __clamp_int(cfg?.n ?? 5, 2, 32)
const sampler_opts = __sampler_opts_for(cfg)
const run_parallel = cfg?.parallel ?? true
const base_opts = if type_of(cfg?.call_opts) == "dict" {
cfg.call_opts
} else {
{}
}
const samples = __collect_samples(prompt, system, base_opts, sampler_opts, n, run_parallel)
const valid = samples.filter({ s -> s?.ok ?? false })
if len(valid) == 0 {
return {ok: false, status: "all_samples_failed", candidates: samples}
}
const reward_filtered = __reward_filter(samples, cfg?.reward)
const judge_kind = cfg?.judge ?? "structured"
if type_of(judge_kind) == "closure" || type_of(judge_kind) == "function" {
const r = try {
judge_kind(valid)
}
if is_err(r) {
return {ok: false, status: "exception", error: unwrap_err(r), candidates: samples}
}
const verdict = unwrap(r)
const chosen_idx = to_int(verdict?.best_index ?? 0)
const safe_idx = if chosen_idx >= 0 && chosen_idx < len(samples) {
chosen_idx
} else {
valid[0].index
}
const best = samples[safe_idx]
return {
ok: true,
best: {text: best?.text ?? "", index: safe_idx, score: verdict?.score ?? nil},
candidates: samples,
judge: "closure",
reasoning: to_string(verdict?.reasoning ?? ""),
}
}
const verdict = __structured_judge(samples, cfg)
if !(verdict.ok ?? false) {
return {
ok: false,
status: to_string(verdict?.status ?? "judge_failed"),
candidates: samples,
reward_filtered: reward_filtered,
}
}
const chosen_idx = verdict.best_index
const safe_idx = if chosen_idx >= 0 && chosen_idx < len(samples) {
chosen_idx
} else {
valid[0].index
}
const best = samples[safe_idx]
return {
ok: true,
best: {text: best?.text ?? "", index: safe_idx},
candidates: samples,
judge: "structured",
reasoning: verdict.reasoning,
scores: verdict.scores,
}
}
// -------------------------------------------------------------------------------------------------
// self_consistency
// -------------------------------------------------------------------------------------------------
fn __log2(x) {
// log2(x) = ln(x) / ln(2). Harn provides `ln`/`log` via std/math; fall
// back on a manual approximation if not.
const lnx = ln(x)
const ln2 = ln(2.0)
return lnx / ln2
}
fn __shannon_entropy(counts, total) {
if total <= 0 {
return 0.0
}
let h = 0.0
for c in counts {
const cnt = to_float(c)
if cnt > 0.0 {
const p = cnt / to_float(total)
h = h - p * __log2(p)
}
}
return h
}
fn __answer_key(value) {
const kind = type_of(value)
if kind == "string" {
return value
}
if kind == "dict" || kind == "list" {
return json_stringify(value)
}
return to_string(value)
}
fn __extract_answer(sample, extract, normalize) {
if !(sample?.ok ?? false) {
return nil
}
const raw = try {
extract(sample.text)
}
if is_err(raw) {
return nil
}
let ans = unwrap(raw)
if normalize != nil && ans != nil {
const norm = try {
normalize(ans)
}
if !is_err(norm) {
ans = unwrap(norm)
}
}
return ans
}
fn __weight_for(sample, vote_mode, confidence_fn) {
if vote_mode != "weighted" {
return 0.0
}
const cv = try {
confidence_fn(sample)
}
if is_err(cv) {
return 1.0
}
return to_float(unwrap(cv))
}
fn __tally_samples(samples, extract, normalize, vote_mode, confidence_fn) {
let keys_in_order = []
let counts = {}
let first_seen = {}
let weights = {}
let extracted = []
let idx = 0
for s in samples {
const ans = __extract_answer(s, extract, normalize)
if ans == nil {
extracted = extracted.push(nil)
} else {
const key = __answer_key(ans)
if counts[key] == nil {
keys_in_order = keys_in_order.push(key)
counts = counts + {[key]: 0}
first_seen = first_seen + {[key]: idx}
weights = weights + {[key]: 0.0}
}
counts = counts + {[key]: counts[key] + 1}
const w = __weight_for(s, vote_mode, confidence_fn)
if w != 0.0 {
weights = weights + {[key]: weights[key] + w}
}
extracted = extracted.push(ans)
}
idx = idx + 1
}
return {
keys_in_order: keys_in_order,
counts: counts,
first_seen: first_seen,
weights: weights,
extracted: extracted,
}
}
fn __score_for(key, counts, weights, vote_mode) {
if vote_mode == "weighted" {
return weights[key]
}
return to_float(counts[key])
}
fn __pick_winner(tally, vote_mode) {
const keys = tally.keys_in_order
let best_key = keys[0]
let best_score = __score_for(best_key, tally.counts, tally.weights, vote_mode)
let best_first_seen = tally.first_seen[best_key]
let tie = false
let i = 1
while i < len(keys) {
const k = keys[i]
const score = __score_for(k, tally.counts, tally.weights, vote_mode)
if score > best_score {
best_key = k
best_score = score
best_first_seen = tally.first_seen[k]
tie = false
} else if score == best_score {
const fs = tally.first_seen[k]
if fs < best_first_seen {
best_key = k
best_first_seen = fs
}
tie = true
}
i = i + 1
}
return {key: best_key, tie: tie}
}
fn __canonical_answer(extracted, key) {
for e in extracted {
if e != nil && __answer_key(e) == key {
return e
}
}
return nil
}
fn __build_distribution(keys_in_order, counts) {
let distribution = []
let total = 0
for k in keys_in_order {
const c = counts[k]
total = total + c
distribution = distribution.push({answer: k, count: c})
}
return {distribution: distribution, total: total}
}
/**
* self_consistency(prompt, system, opts) -> dict
*
* Sample N times at temperature > 0, extract a canonical answer per sample,
* and majority-vote (Wang et al. 2022; arxiv:2203.11171).
*
* Options:
* - n: int (default 8)
* - sampler_opts: dict (default {temperature: 1.2})
* - extract: closure(text) -> any (REQUIRED)
* - normalize: closure(answer) -> any (default identity)
* - vote: "majority" | "weighted" (default "majority")
* - confidence_fn: closure(sample) -> float (REQUIRED for "weighted")
* - parallel: bool (default true)
* - _session_id: string (optional; emits "self_consistency_tie")
*
* Returns: {ok, answer, answer_count, total, distribution, candidates, entropy}.
* Ties: lowest-index extracted answer wins; emits "self_consistency_tie".
* No extractable answers: {ok: false, status: "no_extractable_answers"}.
*
* @effects: []
* @errors: []
*/
pub fn self_consistency(prompt, system, opts = nil) {
const cfg = if type_of(opts) == "dict" {
opts
} else {
{}
}
const extract = cfg?.extract
if extract == nil {
throw "self_consistency: opts.extract is required"
}
const vote_mode = cfg?.vote ?? "majority"
const confidence_fn = cfg?.confidence_fn
if vote_mode == "weighted" && confidence_fn == nil {
throw "self_consistency: opts.confidence_fn is required for weighted voting"
}
const n = __clamp_int(cfg?.n ?? 8, 2, 64)
const sampler_opts = if type_of(cfg?.sampler_opts) == "dict" {
cfg.sampler_opts
} else {
{temperature: 1.2}
}
const run_parallel = cfg?.parallel ?? true
const base_opts = if type_of(cfg?.call_opts) == "dict" {
cfg.call_opts
} else {
{}
}
const samples = __collect_samples(prompt, system, base_opts, sampler_opts, n, run_parallel)
const tally = __tally_samples(samples, extract, cfg?.normalize, vote_mode, confidence_fn)
if len(tally.keys_in_order) == 0 {
return {ok: false, status: "no_extractable_answers", candidates: samples}
}
const winner = __pick_winner(tally, vote_mode)
const dist = __build_distribution(tally.keys_in_order, tally.counts)
let counts_only = []
for d in dist.distribution {
counts_only = counts_only.push(d.count)
}
const session_id = to_string(cfg?._session_id ?? "")
if winner.tie && session_id != "" {
const _ = try {
agent_emit_event(
session_id,
"self_consistency_tie",
{answer: winner.key, total: dist.total, distribution: dist.distribution},
)
}
}
return {
ok: true,
answer: __canonical_answer(tally.extracted, winner.key),
answer_count: tally.counts[winner.key],
total: dist.total,
distribution: dist.distribution,
candidates: samples,
entropy: __shannon_entropy(counts_only, dist.total),
}
}
// -------------------------------------------------------------------------------------------------
// parallel_judge
// -------------------------------------------------------------------------------------------------
/**
* parallel_judge(items, judge_fn, opts) -> list<dict>
*
* Run `judge_fn(item) -> verdict` for each item. Output preserves input
* order. Each entry: {item, verdict?, ok, error?, duration_ms}.
*
* Options:
* - max_concurrent: int (default 4)
* - on_error: "skip" | "fail_fast" | "collect" (default "collect")
* - on_progress: closure({index, total, item, verdict}) -> nil
*
* @effects: []
* @errors: []
*/
pub fn parallel_judge(items, judge_fn, opts = nil) {
const cfg = if type_of(opts) == "dict" {
opts
} else {
{}
}
const on_error = cfg?.on_error ?? "collect"
const on_progress = cfg?.on_progress
const total = len(items)
if total == 0 {
return []
}
// index pairs to preserve order through parallel each
let indexed = []
let i = 0
for item in items {
indexed = indexed.push({index: i, item: item})
i = i + 1
}
const raw = parallel each indexed with { max_concurrent: __clamp_int(cfg?.max_concurrent ?? 4, 1, 64) } { entry ->
let started = harness.clock.now_ms()
let r = try {
judge_fn(entry.item)
}
let duration = harness.clock.now_ms() - started
if is_err(r) {
let err = unwrap_err(r)
let record = {index: entry.index, item: entry.item, ok: false, error: err, duration_ms: duration}
if on_progress != nil {
let _ = try {
on_progress({index: entry.index, total: total, item: entry.item, verdict: nil})
}
}
record
} else {
let verdict = unwrap(r)
let record = {index: entry.index, item: entry.item, ok: true, verdict: verdict, duration_ms: duration}
if on_progress != nil {
let _ = try {
on_progress({index: entry.index, total: total, item: entry.item, verdict: verdict})
}
}
record
}
}
// Sort by index to preserve input order. parallel each currently returns
// results in input-list order, but we sort defensively in case that ever
// changes.
const sorted = raw.sort_by({ entry -> entry.index })
let out = []
for r in sorted {
const failed = !(r?.ok ?? false)
if failed && on_error == "fail_fast" {
throw r?.error ?? "parallel_judge: judge_fn errored"
}
if !(failed && on_error == "skip") {
out = out.push(r)
}
}
return out
}
type DebateResponse = {
debater: string,
text: string,
provider: string?,
model: string?,
input_tokens: int?,
output_tokens: int?,
}
type DebateRound = {
round: int,
responses: list<DebateResponse>,
max_round_drift: float?,
stable: bool?,
drifts: list<dict>?,
}
type DebateResult = {
prompt: string,
debaters: list<string>,
requested_rounds: int,
n_rounds: int,
completed_rounds: int,
stopped_early: bool,
stop_reason: string?,
rounds: list<DebateRound>,
stability: list<dict>,
final_responses: list<DebateResponse>,
short_circuit_event_id: int?,
}
fn __debate_prompt_value(opts) {
const prompt = opts?.prompt ?? opts?.question ?? opts?.task
if type_of(prompt) != "string" || trim(prompt) == "" {
throw "debate: opts.prompt must be a non-empty string"
}
return prompt
}
fn __debate_int(value, fallback, label) {
const parsed = to_int(value ?? fallback)
if parsed == nil || parsed < 1 {
throw "debate: " + label + " must be a positive integer"
}
return parsed
}
fn __debate_float(value, fallback, label) {
const parsed = to_float(value ?? fallback)
if parsed == nil {
throw "debate: " + label + " must be numeric"
}
return parsed
}
fn __debate_debaters(opts) {
const debaters = opts?.debaters ?? opts?.agents ?? ["debater_1", "debater_2"]
if type_of(debaters) != "list" || len(debaters) == 0 {
throw "debate: opts.debaters must be a non-empty list"
}
return debaters
}
fn __debater_name(debater, index) {
if type_of(debater) == "string" {
const name = trim(debater)
if name != "" {
return name
}
}
if type_of(debater) == "dict" {
const name = debater?.name
if type_of(name) == "string" && trim(name) != "" {
return trim(name)
}
return "debater_" + to_string(index + 1)
}
throw "debate: each debater must be a string or dict"
}
fn __debater_instruction(debater) {
if type_of(debater) == "dict" && type_of(debater?.instruction) == "string" {
return trim(debater.instruction)
}
return ""
}
fn __debater_names(debaters) {
let names = []
let index = 0
while index < len(debaters) {
names = names.push(__debater_name(debaters[index], index))
index += 1
}
return names
}
fn __debate_options(raw) {
if type_of(raw) != "dict" {
throw "debate: opts must be a dict"
}
if contains(raw.keys(), "adaptive_stop") && type_of(raw.adaptive_stop) != "bool" {
throw "debate: opts.adaptive_stop must be a bool"
}
const threshold = __debate_float(raw?.stability_threshold ?? raw?.threshold, 0.15, "stability_threshold")
if threshold <= 0.0 || threshold > 1.0 {
throw "debate: stability_threshold must be > 0 and <= 1"
}
return raw
+ {
prompt: __debate_prompt_value(raw),
debaters: __debate_debaters(raw),
n_rounds: __debate_int(raw?.n_rounds ?? raw?.rounds ?? raw?.max_rounds, 3, "n_rounds"),
adaptive_stop: raw?.adaptive_stop ?? false,
stability_threshold: threshold,
stability_patience: 2,
}
}
fn __copy_top_level_llm_options(opts, base) {
let out = {} + base
for key in [
"provider",
"model",
"temperature",
"max_tokens",
"timeout_ms",
"budget",
"response_format",
"schema_retries",
"session_id",
] {
if contains(opts.keys(), key) {
out[key] = opts[key]
}
}
return out
}
fn __debate_llm_options(opts, debater) {
let out = opts?.llm_options ?? {}
if type_of(out) != "dict" {
throw "debate: opts.llm_options must be a dict when provided"
}
out = __copy_top_level_llm_options(opts, out)
if type_of(debater) == "dict" {
const debater_options = debater?.llm_options ?? {}
if type_of(debater_options) != "dict" {
throw "debate: debater.llm_options must be a dict when provided"
}
out = out + debater_options
}
return out
}
fn __debate_system(opts, debater) {
let parts = []
if type_of(opts?.system) == "string" && trim(opts.system) != "" {
parts = parts.push(trim(opts.system))
}
if type_of(debater) == "dict" && type_of(debater?.system) == "string" && trim(debater.system) != "" {
parts = parts.push(trim(debater.system))
}
parts = parts
.push(
"You are participating in a multi-agent debate. Answer from your assigned perspective, revise when prior rounds change your view, and keep the response concise.",
)
return join(parts, "\n\n")
}
fn __debate_history(rounds) {
let lines = []
for round in rounds {
lines = lines.push("Round " + to_string(round.round) + ":")
for response in round.responses {
lines = lines.push(response.debater + ": " + response.text)
}
}
return join(lines, "\n")
}
fn __debate_call_prompt(opts, debater, debater_name, round_number, previous_rounds) {
let parts = [
"Question:\n" + opts.prompt,
"Debater: " + debater_name,
"Round: " + to_string(round_number) + " of " + to_string(opts.n_rounds),
]
const instruction = __debater_instruction(debater)
if instruction != "" {
parts = parts.push("Perspective:\n" + instruction)
}
const history = __debate_history(previous_rounds)
if history != "" {
parts = parts.push("Previous rounds:\n" + history)
}
parts = parts.push("Return only this debater's next response.")
return join(parts, "\n\n")
}
fn __debate_response_text(result) {
if type_of(result) == "dict" {
if result?.text != nil {
return to_string(result.text)
}
if result?.data != nil {
return to_string(result.data)
}
}
return to_string(result)
}
fn __debate_response(debater_name, result) {
let response = {debater: debater_name, text: __debate_response_text(result)}
if type_of(result) == "dict" {
for key in ["provider", "model", "input_tokens", "output_tokens"] {
if result[key] != nil {
response[key] = result[key]
}
}
}
return response
}
fn __debate_run_round(opts, round_number, previous_rounds) {
let responses = []
let index = 0
while index < len(opts.debaters) {
const debater = opts.debaters[index]
const name = __debater_name(debater, index)
const result = llm_call(
__debate_call_prompt(opts, debater, name, round_number, previous_rounds),
__debate_system(opts, debater),
__debate_llm_options(opts, debater),
)
responses = responses.push(__debate_response(name, result))
index += 1
}
return responses
}
fn __debate_tokens(text) {
const cleaned = trim(regex_replace("[^A-Za-z0-9_]+", " ", lowercase(to_string(text))))
if cleaned == "" {
return []
}
return split(cleaned, " ").filter({ token -> return token != "" })
}
fn __ngram_counts(tokens, width) {
let counts = {}
if len(tokens) < width {
return counts
}
let index = 0
while index <= len(tokens) - width {
const key = join(tokens[index:index + width], " ")
const current = counts[key] ?? 0
counts[key] = current + 1
index += 1
}
return counts
}
fn __ngram_precision(reference_tokens, candidate_tokens, width) {
const candidate_counts = __ngram_counts(candidate_tokens, width)
const reference_counts = __ngram_counts(reference_tokens, width)
let overlap = 0
let total = 0
for entry in entries(candidate_counts) {
overlap += min(entry.value, reference_counts[entry.key] ?? 0)
total += entry.value
}
if total == 0 {
return 0.0
}
return overlap * 1.0 / total
}
fn __bleu_lite(reference_text, candidate_text) {
const reference_tokens = __debate_tokens(reference_text)
const candidate_tokens = __debate_tokens(candidate_text)
if len(reference_tokens) == 0 && len(candidate_tokens) == 0 {
return 1.0
}
if len(reference_tokens) == 0 || len(candidate_tokens) == 0 {
return 0.0
}
const unigram = __ngram_precision(reference_tokens, candidate_tokens, 1)
let precision_sum = unigram
let precision_count = 1.0
if len(reference_tokens) >= 2 && len(candidate_tokens) >= 2 {
precision_sum += __ngram_precision(reference_tokens, candidate_tokens, 2)
precision_count += 1.0
}
const brevity = if len(candidate_tokens) < len(reference_tokens) {
len(candidate_tokens) * 1.0 / len(reference_tokens)
} else {
1.0
}
return brevity * precision_sum / precision_count
}
fn __text_drift(previous_text, current_text) {
const similarity = __bleu_lite(previous_text, current_text)
if similarity < 0.0 {
return 1.0
}
if similarity > 1.0 {
return 0.0
}
return 1.0 - similarity
}
fn __round_stability(previous_round, current_round, threshold) {
let drifts = []
let max_drift = 0.0
let index = 0
while index < len(current_round.responses) {
const current = current_round.responses[index]
const previous = previous_round.responses[index]
const drift = __text_drift(previous?.text ?? "", current?.text ?? "")
drifts = drifts.push({debater: current.debater, drift: drift})
max_drift = max(max_drift, drift)
index += 1
}
return {
round: current_round.round,
max_round_drift: max_drift,
threshold: threshold,
stable: max_drift < threshold,
drifts: drifts,
}
}
fn __emit_stability_short_circuit(opts, stability) {
return event_log
.emit(
"llm.ensemble.debate",
"debate_stability_short_circuit",
{
round: stability.round,
requested_rounds: opts.n_rounds,
max_round_drift: stability.max_round_drift,
threshold: opts.stability_threshold,
consecutive_stable_rounds: opts.stability_patience,
drifts: stability.drifts,
},
{round: to_string(stability.round), requested_rounds: to_string(opts.n_rounds)},
)
}
/**
* Run a multi-debater LLM debate. Set `adaptive_stop: true` to stop after
* two consecutive stable rounds when every debater's response drift is below
* `stability_threshold` (default `0.15`).
*
* @effects: []
* @errors: []
*/
pub fn debate(opts) -> DebateResult {
const config = __debate_options(opts)
let rounds = []
let stability = []
let consecutive_stable = 0
let short_circuit_event_id = nil
let round_number = 1
while round_number <= config.n_rounds {
const responses = __debate_run_round(config, round_number, rounds)
let round = {round: round_number, responses: responses}
if len(rounds) > 0 {
const round_stability = __round_stability(rounds[-1], round, config.stability_threshold)
stability = stability.push(round_stability)
round = round
+ {
max_round_drift: round_stability.max_round_drift,
stable: round_stability.stable,
drifts: round_stability.drifts,
}
if config.adaptive_stop {
if round_stability.stable {
consecutive_stable += 1
} else {
consecutive_stable = 0
}
}
}
rounds = rounds.push(round)
if config.adaptive_stop && consecutive_stable >= config.stability_patience {
short_circuit_event_id = __emit_stability_short_circuit(config, stability[-1])
break
}
round_number += 1
}
const stopped_early = short_circuit_event_id != nil
return {
prompt: config.prompt,
debaters: __debater_names(config.debaters),
requested_rounds: config.n_rounds,
n_rounds: config.n_rounds,
completed_rounds: len(rounds),
stopped_early: stopped_early,
stop_reason: if stopped_early {
"stability"
} else {
nil
},
rounds: rounds,
stability: stability,
final_responses: rounds[-1].responses,
short_circuit_event_id: short_circuit_event_id,
}
}
fn __tot_is_callable(value) {
const kind = type_of(value)
return kind == "function" || kind == "closure" || kind == "fn"
}
fn __tot_is_number(value) {
const kind = type_of(value)
return kind == "int" || kind == "float"
}
fn __tot_require_callable(opts, name) {
const value = opts[name]
require __tot_is_callable(value), "tree_of_thoughts: opts." + name + " must be callable"
return value
}
fn __tot_require_positive_int(value, label) {
require __tot_is_number(value), "tree_of_thoughts: " + label + " must be numeric"
require floor(value) == value, "tree_of_thoughts: " + label + " must be an integer"
require value >= 1, "tree_of_thoughts: " + label + " must be at least 1"
return to_int(value)
}
fn __tot_require_non_negative_int(value, label) {
require __tot_is_number(value), "tree_of_thoughts: " + label + " must be numeric"
require floor(value) == value, "tree_of_thoughts: " + label + " must be an integer"
require value >= 0, "tree_of_thoughts: " + label + " must be non-negative"
return to_int(value)
}
fn __tot_score(evaluate, state) {
const score = evaluate(state)
require __tot_is_number(score), "tree_of_thoughts: evaluate(state) must return a number"
return score
}
fn __tot_terminal(is_terminal, state) {
const terminal = is_terminal(state)
require type_of(terminal) == "bool", "tree_of_thoughts: is_terminal(state) must return a bool"
return terminal
}
fn __tot_stop(stop, node) {
if stop == nil {
return false
}
const should_stop = stop(node)
require type_of(should_stop) == "bool", "tree_of_thoughts: stop(node) must return a bool"
return should_stop
}
fn __tot_node(id, parent_id, state, path, depth, score, terminal) {
return {
id: id,
parent_id: parent_id,
state: state,
path: path,
depth: depth,
score: score,
terminal: terminal,
}
}
fn __tot_better_terminal(best_terminal, candidate) {
return candidate.terminal && (best_terminal == nil || candidate.score > best_terminal.score)
}
fn __tot_better_seen(best_seen, candidate) {
return best_seen == nil || candidate.score > best_seen.score
}
fn __tot_update_best(best, node) {
let best_terminal = best.terminal
let best_seen = best.seen
if __tot_better_terminal(best_terminal, node) {
best_terminal = node
}
if __tot_better_seen(best_seen, node) {
best_seen = node
}
return {terminal: best_terminal, seen: best_seen}
}
fn __tot_top(nodes, limit) {
return nodes.sort_by({ node -> 0 - node.score }).take(limit)
}
fn __tot_children(parent, expand, evaluate, is_terminal, k, next_id) {
const expanded = expand(parent.state, k)
require type_of(expanded) == "list", "tree_of_thoughts: expand(state, k) must return a list"
let children = []
let idx = 0
const limit = if len(expanded) < k {
len(expanded)
} else {
k
}
while idx < limit {
const state = expanded[idx]
const node = __tot_node(
next_id + idx,
parent.id,
state,
parent.path.push(state),
parent.depth + 1,
__tot_score(evaluate, state),
__tot_terminal(is_terminal, state),
)
children = children.push(node)
idx = idx + 1
}
return children
}
fn __tot_finalize(search, k, max_depth, beam_width, nodes, best, stop_reason) {
const best_node = best.terminal ?? best.seen
const reason = if stop_reason != nil {
stop_reason
} else if best.terminal != nil {
"terminal"
} else {
"exhausted"
}
return {
ok: best.terminal != nil,
best_path: best_node?.path ?? [],
best_score: best_node?.score,
best_node: best_node,
stop_reason: reason,
tree: {
root_id: 0,
best_id: best_node?.id,
nodes: nodes,
search: search,
k: k,
max_depth: max_depth,
beam_width: beam_width,
},
}
}
fn __tot_search_frontier(
search,
root,
expand,
evaluate,
is_terminal,
stop,
k,
max_depth,
beam_width,
) {
let nodes = [root]
let frontier = [root]
let next_id = 1
let best = __tot_update_best({terminal: nil, seen: nil}, root)
let stop_reason = nil
while len(frontier) > 0 && stop_reason == nil {
const current = if search == "dfs" {
frontier.last()
} else {
frontier[0]
}
frontier = if search == "dfs" {
frontier.pop()
} else {
frontier.slice(1)
}
if __tot_stop(stop, current) {
stop_reason = "stop"
} else if !current.terminal && current.depth < max_depth {
const children = __tot_children(current, expand, evaluate, is_terminal, k, next_id)
next_id = next_id + len(children)
for child in children {
nodes = nodes.push(child)
best = __tot_update_best(best, child)
}
if search == "bfs" {
for child in children {
frontier = frontier.push(child)
}
}
if search == "dfs" {
frontier = frontier + children.reverse()
}
}
}
return __tot_finalize(search, k, max_depth, beam_width, nodes, best, stop_reason)
}
fn __tot_search_beam(root, expand, evaluate, is_terminal, stop, k, max_depth, beam_width) {
let nodes = [root]
let beam = [root]
let next_id = 1
let best = __tot_update_best({terminal: nil, seen: nil}, root)
let stop_reason = nil
let depth = 0
while len(beam) > 0 && depth < max_depth && stop_reason == nil {
let candidates = []
for current in beam {
if stop_reason == nil {
if __tot_stop(stop, current) {
stop_reason = "stop"
} else if !current.terminal && current.depth < max_depth {
const children = __tot_children(current, expand, evaluate, is_terminal, k, next_id)
next_id = next_id + len(children)
for child in children {
nodes = nodes.push(child)
candidates = candidates.push(child)
best = __tot_update_best(best, child)
}
}
}
}
beam = __tot_top(candidates, beam_width)
depth = depth + 1
}
return __tot_finalize("beam", k, max_depth, beam_width, nodes, best, stop_reason)
}
/**
* tree_of_thoughts searches explicitly expanded reasoning states with BFS, DFS, or beam search.
*
* @effects: []
* @errors: []
*/
pub fn tree_of_thoughts(opts) {
require type_of(opts) == "dict", "tree_of_thoughts: opts must be a dict"
require opts.has("initial_state"), "tree_of_thoughts: opts.initial_state is required"
const expand = __tot_require_callable(opts, "expand")
const evaluate = __tot_require_callable(opts, "evaluate")
const is_terminal = __tot_require_callable(opts, "is_terminal")
const stop = opts?.stop
if stop != nil {
require __tot_is_callable(stop), "tree_of_thoughts: opts.stop must be callable"
}
const search = opts?.search ?? "bfs"
require search == "bfs" || search == "dfs" || search == "beam", "tree_of_thoughts: opts.search must be \"bfs\", \"dfs\", or \"beam\""
const k = __tot_require_positive_int(opts?.k ?? 1, "opts.k")
const max_depth = __tot_require_non_negative_int(opts?.max_depth ?? 8, "opts.max_depth")
const beam_width = __tot_require_positive_int(opts?.beam_width ?? k, "opts.beam_width")
const initial_state = opts.initial_state
const root = __tot_node(
0,
nil,
initial_state,
[initial_state],
0,
__tot_score(evaluate, initial_state),
__tot_terminal(is_terminal, initial_state),
)
if search == "beam" {
return __tot_search_beam(root, expand, evaluate, is_terminal, stop, k, max_depth, beam_width)
}
return __tot_search_frontier(search, root, expand, evaluate, is_terminal, stop, k, max_depth, beam_width)
}