be65399710
The /score endpoint was receiving chat-format messages which had to go through the chat template tokenizer — this was failing with "System message must be first" errors because the AST structure doesn't map cleanly to chat message format. Send raw token IDs via the new `prompt` field instead, matching what the /completions endpoint already does. The vLLM score endpoint finds assistant boundaries by scanning for <|im_start|>assistant token patterns, so no message-level metadata is needed. Also includes identity and journal sections in the scored context, matching what the model actually sees during inference. Co-Authored-By: Proof of Concept <poc@bcachefs.org>
440 lines
15 KiB
Rust
440 lines
15 KiB
Rust
// training.rs — Memory importance scoring via /v1/score
|
||
//
|
||
// Three scoring modes, all built on the same call_score() primitive:
|
||
//
|
||
// score_memories() — Full N×M matrix (memories × responses) for the
|
||
// debug screen. Expensive: N+1 API calls.
|
||
//
|
||
// memory_score() — Single memory importance. Scores the 50 messages
|
||
// after it was surfaced, with/without that memory.
|
||
// 2 API calls.
|
||
//
|
||
// finetune_score() — Identifies training candidates. Scores recent
|
||
// messages with all memories stripped. Responses
|
||
// with high divergence depend on memories the model
|
||
// hasn't internalized. 2 API calls.
|
||
|
||
use crate::agent::api::ApiClient;
|
||
use crate::agent::context::{AstNode, Ast, NodeBody, ContextState, Role};
|
||
|
||
const SCORE_TIMEOUT: std::time::Duration = std::time::Duration::from_secs(120);
|
||
|
||
// ── Message building ────────────────────────────────────────────
|
||
|
||
/// What to filter when building the message array for scoring.
|
||
#[allow(dead_code)]
|
||
enum Filter<'a> {
|
||
None,
|
||
SkipIndex(usize),
|
||
SkipKey(&'a str),
|
||
SkipAllMemories,
|
||
}
|
||
|
||
fn is_memory(node: &AstNode) -> bool {
|
||
matches!(node, AstNode::Leaf(leaf) if matches!(leaf.body(), NodeBody::Memory { .. }))
|
||
}
|
||
|
||
fn memory_key(node: &AstNode) -> Option<&str> {
|
||
match node {
|
||
AstNode::Leaf(leaf) => match leaf.body() {
|
||
NodeBody::Memory { key, .. } => Some(key),
|
||
_ => None,
|
||
},
|
||
_ => None,
|
||
}
|
||
}
|
||
|
||
fn is_assistant(node: &AstNode) -> bool {
|
||
matches!(node, AstNode::Branch { role: Role::Assistant, .. })
|
||
}
|
||
|
||
/// Build a token ID array for a scoring call.
|
||
///
|
||
/// Includes all sections up to and including conversation entries in
|
||
/// `range`, with `filter` applied to conversation entries.
|
||
fn build_token_ids(
|
||
context: &ContextState,
|
||
range: std::ops::Range<usize>,
|
||
filter: Filter,
|
||
) -> Vec<u32> {
|
||
use crate::agent::context::Ast;
|
||
let mut ids = Vec::new();
|
||
for node in context.system() {
|
||
ids.extend(node.token_ids());
|
||
}
|
||
for node in context.identity() {
|
||
ids.extend(node.token_ids());
|
||
}
|
||
for node in context.journal() {
|
||
ids.extend(node.token_ids());
|
||
}
|
||
let entries = context.conversation();
|
||
for i in range {
|
||
let node = &entries[i];
|
||
let skip = match &filter {
|
||
Filter::None => false,
|
||
Filter::SkipIndex(idx) => i == *idx,
|
||
Filter::SkipKey(key) => memory_key(node) == Some(*key),
|
||
Filter::SkipAllMemories => is_memory(node),
|
||
};
|
||
if skip { continue; }
|
||
ids.extend(node.token_ids());
|
||
}
|
||
ids
|
||
}
|
||
|
||
// ── Score API ───────────────────────────────────────────────────
|
||
|
||
#[derive(serde::Deserialize)]
|
||
struct ScoreResult {
|
||
total_logprob: f64,
|
||
}
|
||
|
||
#[derive(serde::Deserialize)]
|
||
struct ScoreResponse {
|
||
scores: Vec<ScoreResult>,
|
||
}
|
||
|
||
fn http_client() -> crate::agent::api::http::HttpClient {
|
||
crate::agent::api::http::HttpClient::builder()
|
||
.timeout(SCORE_TIMEOUT)
|
||
.build()
|
||
}
|
||
|
||
async fn call_score(
|
||
http: &crate::agent::api::http::HttpClient,
|
||
client: &ApiClient,
|
||
prompt: &[u32],
|
||
priority: Option<i32>,
|
||
) -> anyhow::Result<Vec<ScoreResult>> {
|
||
let url = format!("{}/score", client.base_url());
|
||
let auth = format!("Bearer {}", client.api_key());
|
||
let mut body = serde_json::json!({
|
||
"model": client.model,
|
||
"prompt": prompt,
|
||
"logprobs": 1,
|
||
});
|
||
if let Some(p) = priority {
|
||
body["priority"] = serde_json::json!(p);
|
||
}
|
||
let response = http
|
||
.send_json("POST", &url, &[
|
||
("authorization", &auth),
|
||
], &body)
|
||
.await?;
|
||
|
||
let status = response.status();
|
||
let body: serde_json::Value = response.json().await?;
|
||
|
||
if !status.is_success() {
|
||
let msg = body.get("error").and_then(|e| e.as_str()).unwrap_or("unknown error");
|
||
anyhow::bail!("score API HTTP {}: {}", status, msg);
|
||
}
|
||
if let Some(err) = body.get("error").and_then(|e| e.as_str()) {
|
||
anyhow::bail!("score API error: {}", err);
|
||
}
|
||
|
||
let result: ScoreResponse = serde_json::from_value(body)
|
||
.map_err(|e| anyhow::anyhow!("failed to parse score response: {}", e))?;
|
||
Ok(result.scores)
|
||
}
|
||
|
||
/// Compute per-position logprob divergence: how much worse the model
|
||
/// scores each response without something vs with it.
|
||
fn divergence(baseline: &[ScoreResult], without: &[ScoreResult]) -> Vec<f64> {
|
||
baseline.iter().enumerate()
|
||
.map(|(i, base)| {
|
||
let without_lp = without.get(i).map(|s| s.total_logprob).unwrap_or(base.total_logprob);
|
||
(base.total_logprob - without_lp).max(0.0)
|
||
})
|
||
.collect()
|
||
}
|
||
|
||
/// Score two message sets and return total divergence.
|
||
async fn score_divergence(
|
||
http: &crate::agent::api::http::HttpClient,
|
||
client: &ApiClient,
|
||
context: &ContextState,
|
||
range: std::ops::Range<usize>,
|
||
filter: Filter<'_>,
|
||
priority: Option<i32>,
|
||
) -> anyhow::Result<(Vec<f64>, Vec<ScoreResult>)> {
|
||
let baseline = call_score(http, client, &build_token_ids(context, range.clone(), Filter::None), priority).await?;
|
||
let without = call_score(http, client, &build_token_ids(context, range, filter), priority).await?;
|
||
let divs = divergence(&baseline, &without);
|
||
Ok((divs, baseline))
|
||
}
|
||
|
||
// ── Full matrix scoring (debug screen) ──────────────────────────
|
||
|
||
/// Result of scoring one conversation's memory usage.
|
||
pub struct MemoryScore {
|
||
pub memory_weights: Vec<(String, f64)>,
|
||
pub response_scores: Vec<f64>,
|
||
/// Full matrix: divergence[memory_idx][response_idx]
|
||
pub matrix: Vec<Vec<f64>>,
|
||
pub memory_keys: Vec<String>,
|
||
pub response_entry_indices: Vec<usize>,
|
||
}
|
||
|
||
impl MemoryScore {
|
||
pub fn important_memories_for_entry(&self, entry_idx: usize) -> Vec<(&str, f64)> {
|
||
let Some(resp_idx) = self.response_entry_indices.iter().position(|&i| i == entry_idx)
|
||
else { return Vec::new() };
|
||
|
||
let mut result: Vec<(&str, f64)> = self.memory_keys.iter()
|
||
.zip(self.matrix.iter())
|
||
.filter_map(|(key, row)| {
|
||
let score = row.get(resp_idx).copied().unwrap_or(0.0);
|
||
if score > 0.01 { Some((key.as_str(), score)) } else { None }
|
||
})
|
||
.collect();
|
||
result.sort_by(|a, b| b.1.partial_cmp(&a.1).unwrap_or(std::cmp::Ordering::Equal));
|
||
result
|
||
}
|
||
}
|
||
|
||
/// Score how important each memory is to the conversation (full matrix).
|
||
pub async fn score_memories(
|
||
context: &ContextState,
|
||
client: &ApiClient,
|
||
) -> anyhow::Result<MemoryScore> {
|
||
let mut memory_keys: Vec<String> = context.conversation().iter()
|
||
.filter_map(|node| memory_key(node).map(String::from))
|
||
.collect();
|
||
memory_keys.dedup();
|
||
|
||
let response_indices: Vec<usize> = context.conversation().iter().enumerate()
|
||
.filter(|(_, node)| is_assistant(node))
|
||
.map(|(i, _)| i)
|
||
.collect();
|
||
|
||
if memory_keys.is_empty() || response_indices.is_empty() {
|
||
return Ok(MemoryScore {
|
||
memory_weights: Vec::new(), response_scores: Vec::new(),
|
||
matrix: Vec::new(), memory_keys: Vec::new(),
|
||
response_entry_indices: Vec::new(),
|
||
});
|
||
}
|
||
|
||
|
||
let http = http_client();
|
||
let range = 0..context.conversation().len();
|
||
|
||
let baseline = call_score(&http, client, &build_token_ids(context, range.clone(), Filter::None), Some(5)).await?;
|
||
|
||
let total = memory_keys.len();
|
||
let mut matrix: Vec<Vec<f64>> = Vec::new();
|
||
|
||
for (mem_idx, key) in memory_keys.iter().enumerate() {
|
||
dbglog!(
|
||
"scoring {}/{}: {}...", mem_idx + 1, total, key,
|
||
);
|
||
let msgs = build_token_ids(context, range.clone(), Filter::SkipKey(key));
|
||
match call_score(&http, client, &msgs, Some(5)).await {
|
||
Ok(without) => matrix.push(divergence(&baseline, &without)),
|
||
Err(e) => {
|
||
dbglog!(
|
||
"[training] {} FAILED: {:#}", key, e,
|
||
);
|
||
matrix.push(vec![0.0; baseline.len()]);
|
||
}
|
||
}
|
||
}
|
||
|
||
|
||
let memory_weights: Vec<(String, f64)> = memory_keys.iter()
|
||
.zip(matrix.iter())
|
||
.map(|(key, row)| (key.clone(), row.iter().sum()))
|
||
.collect();
|
||
|
||
let mut response_scores = vec![0.0; response_indices.len()];
|
||
for row in &matrix {
|
||
for (j, &v) in row.iter().enumerate() {
|
||
if j < response_scores.len() { response_scores[j] += v; }
|
||
}
|
||
}
|
||
|
||
Ok(MemoryScore {
|
||
memory_weights, response_scores, matrix, memory_keys,
|
||
response_entry_indices: response_indices,
|
||
})
|
||
}
|
||
|
||
/// Find the entry index after `start` that contains the Nth assistant response.
|
||
/// Returns (end_index, true) if N responses were found, (entries.len(), false) if not.
|
||
fn nth_response_end(entries: &[AstNode], start: usize, n: usize) -> (usize, bool) {
|
||
let mut count = 0;
|
||
for i in start..entries.len() {
|
||
if is_assistant(&entries[i]) {
|
||
count += 1;
|
||
if count >= n { return (i + 1, true); }
|
||
}
|
||
}
|
||
(entries.len(), false)
|
||
}
|
||
|
||
// ── Single memory scoring ───────────────────────────────────────
|
||
|
||
/// Score how important a single memory is to the conversation.
|
||
///
|
||
/// Scores the 50 messages after the memory was surfaced — the window
|
||
/// where it could have influenced responses. Returns the sum of
|
||
/// divergence, or 0.0 if the memory isn't in the conversation.
|
||
pub async fn score_memory(
|
||
context: &ContextState,
|
||
key: &str,
|
||
client: &ApiClient,
|
||
) -> anyhow::Result<f64> {
|
||
const RESPONSE_WINDOW: usize = 50;
|
||
|
||
let entries = context.conversation();
|
||
let first_pos = match entries.iter().position(|node| memory_key(node) == Some(key)) {
|
||
Some(p) => p,
|
||
None => return Ok(0.0),
|
||
};
|
||
|
||
let (end, _) = nth_response_end(entries, first_pos, RESPONSE_WINDOW);
|
||
let range = first_pos..end;
|
||
if !entries[range.clone()].iter().any(|node| is_assistant(node)) {
|
||
return Ok(0.0);
|
||
}
|
||
|
||
let http = http_client();
|
||
let (divs, _) = score_divergence(&http, client, context, range, Filter::SkipKey(key), Some(5)).await?;
|
||
|
||
Ok(divs.iter().sum())
|
||
}
|
||
|
||
// ── Background memory scoring ───────────────────────────────────
|
||
|
||
/// Score memories in the conversation that are due for re-scoring.
|
||
///
|
||
/// Checks the graph for each memory's last_scored timestamp. Scores
|
||
/// nodes that haven't been scored within `max_age_secs`, oldest first.
|
||
/// Updates the graph weight (EWMA) and last_scored after each.
|
||
///
|
||
/// Returns the number of nodes scored and their (key, score) pairs.
|
||
pub async fn score_memories_incremental<F, Fut>(
|
||
context: &ContextState,
|
||
max_age_secs: i64,
|
||
response_window: usize,
|
||
client: &ApiClient,
|
||
agent: &std::sync::Arc<crate::agent::Agent>,
|
||
mut on_score: F,
|
||
) -> anyhow::Result<usize>
|
||
where
|
||
F: FnMut(String, f64) -> Fut,
|
||
Fut: std::future::Future<Output = ()>,
|
||
{
|
||
let now = chrono::Utc::now().timestamp();
|
||
|
||
// Collect unique memory keys with their first position
|
||
let mut seen = std::collections::HashSet::new();
|
||
let mut candidates: Vec<(usize, String, i64)> = Vec::new(); // (pos, key, last_scored)
|
||
|
||
let store = crate::hippocampus::store::Store::load().unwrap_or_default();
|
||
|
||
for (i, node) in context.conversation().iter().enumerate() {
|
||
if let Some(key) = memory_key(node) {
|
||
if !seen.insert(key.to_owned()) { continue; }
|
||
let last_scored = store.nodes.get(key)
|
||
.map(|n| n.last_scored)
|
||
.unwrap_or(0);
|
||
if now - last_scored >= max_age_secs {
|
||
candidates.push((i, key.to_owned(), last_scored));
|
||
}
|
||
}
|
||
}
|
||
|
||
// Score oldest-first
|
||
candidates.sort_by_key(|&(_, _, last)| last);
|
||
|
||
let http = http_client();
|
||
let mut scored = 0;
|
||
|
||
let entries = context.conversation();
|
||
let total_tokens: usize = entries.iter().map(|n| n.tokens()).sum();
|
||
let token_cutoff = total_tokens * 60 / 100;
|
||
|
||
// Precompute cumulative token position for each entry
|
||
let mut cumulative: Vec<usize> = Vec::with_capacity(entries.len());
|
||
let mut running = 0;
|
||
for e in entries {
|
||
running += e.tokens();
|
||
cumulative.push(running);
|
||
}
|
||
|
||
dbglog!("[scoring] total_tokens={}, cutoff={}, {} candidates", total_tokens, token_cutoff, candidates.len());
|
||
|
||
for (pos, key, _) in &candidates {
|
||
// Only score memories in the first 60% of the conversation by tokens —
|
||
// recent memories don't have enough responses to evaluate yet.
|
||
let cum = cumulative.get(*pos).copied().unwrap_or(total_tokens);
|
||
if cum > token_cutoff {
|
||
dbglog!("[scoring] skip {} (tokens {}/{} past cutoff)", key, cum, token_cutoff);
|
||
continue;
|
||
}
|
||
let (end, _) = nth_response_end(context.conversation(), *pos, response_window);
|
||
let range = *pos..end;
|
||
if !context.conversation()[range.clone()].iter().any(|node| is_assistant(node)) {
|
||
dbglog!("[scoring] skip {} (no assistant response in range {}..{})", key, pos, end);
|
||
continue;
|
||
}
|
||
|
||
let _scoring = crate::agent::start_activity(agent, format!("scoring: {}", key)).await;
|
||
match score_divergence(&http, client, context, range, Filter::SkipKey(key), Some(5)).await {
|
||
Ok((divs, _)) => {
|
||
let n_responses = divs.len();
|
||
let max_div = divs.iter().cloned().fold(0.0f64, f64::max);
|
||
dbglog!(
|
||
"[scoring] {} max:{:.3} ({} responses)", key, max_div, n_responses,
|
||
);
|
||
on_score(key.clone(), max_div).await;
|
||
scored += 1;
|
||
}
|
||
Err(e) => {
|
||
dbglog!(
|
||
"[scoring] {} FAILED: {:#}", key, e,
|
||
);
|
||
}
|
||
}
|
||
}
|
||
|
||
Ok(scored)
|
||
}
|
||
|
||
// ── Fine-tuning scoring ─────────────────────────────────────────
|
||
|
||
/// Score which recent responses are candidates for fine-tuning.
|
||
///
|
||
/// Removes all memories and scores the most recent `count` messages.
|
||
/// Responses with high divergence depend on memories the model hasn't
|
||
/// internalized — these are fine-tuning candidates.
|
||
///
|
||
/// Returns (entry_index, divergence) pairs, sorted by divergence descending.
|
||
pub async fn score_finetune(
|
||
context: &ContextState,
|
||
count: usize,
|
||
client: &ApiClient,
|
||
) -> anyhow::Result<Vec<(usize, f64)>> {
|
||
let entries = context.conversation();
|
||
let range = entries.len().saturating_sub(count)..entries.len();
|
||
|
||
let response_positions: Vec<usize> = range.clone()
|
||
.filter(|&i| is_assistant(&entries[i]))
|
||
.collect();
|
||
if response_positions.is_empty() {
|
||
return Ok(Vec::new());
|
||
}
|
||
|
||
let http = http_client();
|
||
let (divs, _) = score_divergence(&http, client, context, range, Filter::SkipAllMemories, Some(5)).await?;
|
||
|
||
let mut results: Vec<(usize, f64)> = response_positions.iter()
|
||
.enumerate()
|
||
.map(|(i, &entry_idx)| (entry_idx, divs.get(i).copied().unwrap_or(0.0)))
|
||
.collect();
|
||
results.sort_by(|a, b| b.1.partial_cmp(&a.1).unwrap_or(std::cmp::Ordering::Equal));
|
||
Ok(results)
|
||
}
|