Deploy Poimen Memory K8s cluster with ArgoCD tracking (M2.2, M3.5-M3.7)

This commit is contained in:
Story Crater Bot
2026-08-22 23:13:42 -07:00
parent af9c5ba01b
commit 695e115212
67 changed files with 8438 additions and 24 deletions
+54
View File
@@ -0,0 +1,54 @@
use anyhow::Result;
use serde::{Deserialize, Serialize};
use std::fs::{create_dir_all, OpenOptions};
use std::io::Write;
use std::path::PathBuf;
/// JSONL event record.
#[derive(Debug, Clone, PartialEq, Serialize, Deserialize)]
pub struct EventRecord {
pub project: String,
pub query: String,
pub run: String,
pub turn: u32,
pub event_type: String,
pub data: serde_json::Value,
}
/// Event log writer.
pub struct LogWriter {
path: PathBuf,
}
impl LogWriter {
/// Open or create log file.
pub fn new(project: &str, query: &str, run: &str) -> Result<Self> {
let dir = PathBuf::from(format!("log/{}/{}", project, query));
create_dir_all(&dir)?;
Ok(Self {
path: dir.join(format!("{}.jsonl", run)),
})
}
/// Append event to log.
pub fn log(&mut self, record: EventRecord) -> Result<()> {
let mut file = OpenOptions::new()
.create(true)
.append(true)
.open(&self.path)?;
serde_json::to_writer(&mut file, &record)?;
file.write_all(b"\n")?;
Ok(())
}
/// Read all events from log.
pub fn read_all(&self) -> Result<Vec<EventRecord>> {
let contents = std::fs::read_to_string(&self.path)?;
contents
.lines()
.filter(|line| !line.is_empty())
.map(|line| serde_json::from_str(line).map_err(|e| anyhow::anyhow!("Parse error: {}", e)))
.collect()
}
}
+11 -1
View File
@@ -1 +1,11 @@
pub mod placeholder {}
pub mod event_log;
pub mod pgvector;
pub mod rebuild;
pub mod pg_repo;
pub mod obsidian;
pub use event_log::{EventRecord, LogWriter};
pub use pgvector::{VectorRecord, VectorStore};
pub use rebuild::RebuildState;
pub use pg_repo::{PgRepo, MemoryNode, VectorKind, Level, ScoredNode};
pub use obsidian::ObsidianProjector;
+193
View File
@@ -0,0 +1,193 @@
use crate::EventRecord;
use anyhow::Result;
use std::collections::{BTreeMap, HashMap};
use std::fs;
/// Obsidian vault projector (deterministic, byte-identical).
pub struct ObsidianProjector {
vault_dir: String,
_emit_evidence: bool,
}
/// Vault note metadata (stable frontmatter order).
#[derive(Debug, Clone)]
pub struct VaultNote {
pub project: String,
pub level: String,
pub query_id: Option<String>,
pub updated: String,
pub chunks_seen: u32,
pub chunks_used: u32,
pub run_id: String,
pub body: String,
pub parents: Vec<(String, String)>,
}
impl ObsidianProjector {
/// Create projector.
pub fn new(_log_dir: &str, vault_dir: &str, emit_evidence: bool) -> Self {
Self {
vault_dir: vault_dir.to_string(),
_emit_evidence: emit_evidence,
}
}
/// Project log to vault (deterministic).
pub fn project(&self, events: &[EventRecord]) -> Result<()> {
fs::create_dir_all(&self.vault_dir)?;
// Group by project and query
let mut by_project: HashMap<String, HashMap<String, Vec<&EventRecord>>> = HashMap::new();
for event in events {
by_project
.entry(event.project.clone())
.or_insert_with(HashMap::new)
.entry(event.query.clone())
.or_insert_with(Vec::new)
.push(event);
}
// Generate notes per project (in sorted order for determinism)
let mut sorted_projects: Vec<_> = by_project.iter().collect();
sorted_projects.sort_by_key(|(p, _)| p.as_str());
for (project, queries) in sorted_projects {
let proj_dir = format!("{}/{}", self.vault_dir, project);
fs::create_dir_all(&proj_dir)?;
// Generate index (L2)
let index_note = VaultNote {
project: project.clone(),
level: "L2".to_string(),
query_id: None,
updated: "2026-01-01".to_string(),
chunks_seen: 0,
chunks_used: 0,
run_id: "index".to_string(),
body: String::new(),
parents: vec![],
};
self.write_note(&proj_dir, "index", &index_note)?;
// Generate per-query notes (L1) in sorted order
let mut sorted_queries: Vec<_> = queries.iter().collect();
sorted_queries.sort_by_key(|(qid, _)| qid.as_str());
for (query_id, query_events) in sorted_queries {
let (chunks_seen, chunks_used, body, parents) =
Self::summarize_query(query_events);
let note = VaultNote {
project: project.clone(),
level: "L1".to_string(),
query_id: Some(query_id.to_string()),
updated: "2026-01-01".to_string(),
chunks_seen,
chunks_used,
run_id: "run1".to_string(),
body,
parents,
};
self.write_note(&proj_dir, query_id, &note)?;
}
}
Ok(())
}
/// Write note with deterministic formatting.
fn write_note(&self, dir: &str, name: &str, note: &VaultNote) -> Result<()> {
// Stable frontmatter order (BTreeMap keeps keys sorted)
let mut fm = BTreeMap::new();
fm.insert("chunks_seen", note.chunks_seen.to_string());
fm.insert("chunks_used", note.chunks_used.to_string());
fm.insert("level", note.level.clone());
fm.insert("project", note.project.clone());
if let Some(qid) = &note.query_id {
fm.insert("query_id", qid.clone());
}
fm.insert("run_id", note.run_id.clone());
fm.insert("updated", note.updated.clone());
// Build frontmatter
let mut content = String::from("---\n");
for (k, v) in fm.iter() {
content.push_str(&format!("{}: {}\n", k, v));
}
content.push_str("---\n");
// Title
let title = note.query_id.as_ref().unwrap_or(&note.project);
content.push_str(&format!("# {}\n\n", title));
// Body
if note.body.is_empty() {
content.push_str("No evidence found.\n\n");
} else {
content.push_str(&note.body);
if !note.body.ends_with('\n') {
content.push('\n');
}
content.push('\n');
}
// Provenance (sorted)
if !note.parents.is_empty() {
content.push_str("## Provenance\n");
let mut sorted_parents = note.parents.clone();
sorted_parents.sort();
for (source, time) in sorted_parents {
content.push_str(&format!("- [[{}-{}]]\n", source, time));
}
}
// Ensure exactly one trailing newline
if !content.ends_with('\n') {
content.push('\n');
}
// Write to file
let path = format!("{}/{}.md", dir, name);
fs::write(&path, &content)?;
Ok(())
}
/// Summarize query events.
fn summarize_query(
events: &[&EventRecord],
) -> (u32, u32, String, Vec<(String, String)>) {
let mut chunks_seen = 0u32;
let mut chunks_used = 0u32;
let mut body = String::new();
let mut parents = Vec::new();
for event in events.iter() {
if event.event_type.contains("Gate") {
chunks_seen += 1;
}
if event.event_type.contains("Evidence") {
chunks_used += 1;
}
// Simplified parent extraction
if let Some(obj) = event.data.as_object() {
if let Some(parent) = obj.get("parent") {
if let Some(s) = parent.as_str() {
parents.push((s.to_string(), format!("t{}", event.turn)));
}
}
}
}
if chunks_used > 0 {
body = format!(
"Extracted from {} chunks, using {}\n",
chunks_seen, chunks_used
);
}
(chunks_seen, chunks_used, body, parents)
}
}
+210
View File
@@ -0,0 +1,210 @@
use anyhow::Result;
use serde::{Deserialize, Serialize};
use std::collections::BTreeMap;
/// Vector kind (text or symptom).
#[derive(Debug, Clone, Copy, PartialEq, Eq, PartialOrd, Ord, Hash, Serialize, Deserialize)]
pub enum VectorKind {
Text,
Symptom,
}
impl std::fmt::Display for VectorKind {
fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
match self {
VectorKind::Text => write!(f, "text"),
VectorKind::Symptom => write!(f, "symptom"),
}
}
}
/// Level (L0, L1, L2).
#[derive(Debug, Clone, Copy, PartialEq, Eq, PartialOrd, Ord, Hash, Serialize, Deserialize)]
pub enum Level {
L0,
L1,
L2,
}
/// Memory node (idempotent upsert key: sha256).
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct MemoryNode {
pub sha256: String,
pub level: Level,
pub project: String,
pub text: String,
pub tokens: u32,
}
/// Scored search result.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct ScoredNode {
pub node: MemoryNode,
pub distance: f32,
pub matched_kind: VectorKind,
}
/// PostgreSQL repository (in-memory mock for now).
pub struct PgRepo {
// Nodes by sha256
nodes: BTreeMap<String, MemoryNode>,
// Vectors by (sha256, kind)
vectors: BTreeMap<(String, VectorKind), Vec<f32>>,
// Parents edges: child_sha -> vec of parent_shas
edges: BTreeMap<String, Vec<String>>,
}
impl PgRepo {
/// Create new repo (mock, no real DB).
pub fn new() -> Self {
Self {
nodes: BTreeMap::new(),
vectors: BTreeMap::new(),
edges: BTreeMap::new(),
}
}
/// Upsert node (idempotent).
pub fn upsert_node(&mut self, node: &MemoryNode) -> Result<()> {
self.nodes.insert(node.sha256.clone(), node.clone());
Ok(())
}
/// Upsert many nodes (batching embedding calls).
pub fn upsert_many(&mut self, nodes: &[MemoryNode]) -> Result<()> {
for node in nodes {
self.upsert_node(node)?;
}
Ok(())
}
/// Upsert vector for node.
pub fn upsert_vector(&mut self, sha: &str, kind: VectorKind, embedding: &[f32]) -> Result<()> {
if !self.nodes.contains_key(sha) {
return Err(anyhow::anyhow!("Node {} not found", sha));
}
self.vectors.insert((sha.to_string(), kind), embedding.to_vec());
Ok(())
}
/// Insert edges (requires both endpoints exist).
pub fn insert_edges(&mut self, child: &str, parents: &[String]) -> Result<()> {
if !self.nodes.contains_key(child) {
return Err(anyhow::anyhow!("Child node {} not found", child));
}
for parent in parents {
if !self.nodes.contains_key(parent) {
return Err(anyhow::anyhow!("Parent node {} not found", parent));
}
}
self.edges.insert(child.to_string(), parents.to_vec());
Ok(())
}
/// Search by cosine distance.
pub fn search(
&self,
q: &[f32],
kind: VectorKind,
levels: &[Level],
) -> Result<Vec<ScoredNode>> {
let mut results = Vec::new();
for ((sha, vkind), embedding) in &self.vectors {
if *vkind != kind {
continue;
}
if let Some(node) = self.nodes.get(sha) {
if !levels.contains(&node.level) {
continue;
}
if let Some(dist) = cosine_distance(q, embedding) {
results.push(ScoredNode {
node: node.clone(),
distance: dist,
matched_kind: kind,
});
}
}
}
// Sort by distance (ascending)
results.sort_by(|a, b| a.distance.partial_cmp(&b.distance).unwrap());
Ok(results)
}
/// Parents of node.
pub fn parents_of(&self, sha: &str) -> Result<Vec<MemoryNode>> {
let parent_shas = self.edges.get(sha).cloned().unwrap_or_default();
let parents: Vec<_> = parent_shas
.iter()
.filter_map(|p_sha| self.nodes.get(p_sha).cloned())
.collect();
Ok(parents)
}
/// Clear all nodes for project.
pub fn clear_project(&mut self, project: &str) -> Result<()> {
let nodes_to_remove: Vec<String> = self
.nodes
.iter()
.filter(|(_, n)| n.project == project)
.map(|(sha, _)| sha.clone())
.collect();
// Remove vectors
self.vectors.retain(|(sha, _), _| !nodes_to_remove.contains(sha));
// Remove edges
self.edges.retain(|child, _| !nodes_to_remove.contains(child));
// Remove nodes
self.nodes.retain(|sha, _| !nodes_to_remove.contains(sha));
Ok(())
}
/// Get all nodes.
pub fn all_nodes(&self) -> Vec<&MemoryNode> {
self.nodes.values().collect()
}
/// Verify: count upserted nodes.
pub fn node_count(&self) -> usize {
self.nodes.len()
}
/// Verify: count edges.
pub fn edge_count(&self) -> usize {
self.edges.len()
}
}
/// Cosine distance (1 - cosine_similarity).
fn cosine_distance(a: &[f32], b: &[f32]) -> Option<f32> {
if a.len() != b.len() || a.is_empty() {
return None;
}
let mut dot = 0.0;
let mut norm_a = 0.0;
let mut norm_b = 0.0;
for (x, y) in a.iter().zip(b.iter()) {
dot += x * y;
norm_a += x * x;
norm_b += y * y;
}
let norm_a = norm_a.sqrt();
let norm_b = norm_b.sqrt();
if norm_a == 0.0 || norm_b == 0.0 {
return None;
}
let similarity = dot / (norm_a * norm_b);
Some(1.0 - similarity)
}
+81
View File
@@ -0,0 +1,81 @@
use anyhow::Result;
use serde::{Deserialize, Serialize};
/// Vector embedding record in pgvector.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct VectorRecord {
pub id: String,
pub chunk_id: String,
pub kind: String, // "text" | "symptom"
pub embedding: Vec<f32>, // 768-dimensional for nomic
pub tokens: u32,
}
/// pgvector client.
pub struct VectorStore {
// In production: PostgreSQL connection
// For now: in-memory vec
records: Vec<VectorRecord>,
}
impl VectorStore {
/// Create a new vector store.
pub fn new() -> Self {
Self {
records: Vec::new(),
}
}
/// Insert a vector record.
pub fn insert(&mut self, record: VectorRecord) -> Result<()> {
self.records.push(record);
Ok(())
}
/// Search by cosine similarity.
pub fn search(&self, query: &[f32], limit: usize, min_score: f32) -> Result<Vec<(String, f32)>> {
let mut results = Vec::new();
for record in &self.records {
if let Some(score) = cosine_similarity(query, &record.embedding) {
if score >= min_score {
results.push((record.id.clone(), score));
}
}
}
results.sort_by(|a, b| b.1.partial_cmp(&a.1).unwrap());
Ok(results.into_iter().take(limit).collect())
}
/// Get all records.
pub fn all(&self) -> Vec<&VectorRecord> {
self.records.iter().collect()
}
}
/// Compute cosine similarity between two vectors.
fn cosine_similarity(a: &[f32], b: &[f32]) -> Option<f32> {
if a.len() != b.len() {
return None;
}
let mut dot_product = 0.0;
let mut norm_a = 0.0;
let mut norm_b = 0.0;
for (x, y) in a.iter().zip(b.iter()) {
dot_product += x * y;
norm_a += x * x;
norm_b += y * y;
}
let norm_a = norm_a.sqrt();
let norm_b = norm_b.sqrt();
if norm_a == 0.0 || norm_b == 0.0 {
return None;
}
Some(dot_product / (norm_a * norm_b))
}
+80
View File
@@ -0,0 +1,80 @@
use crate::EventRecord;
use anyhow::Result;
use serde::{Deserialize, Serialize};
use std::collections::BTreeMap;
/// Deterministic rebuild state from JSONL event log.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct RebuildState {
pub memories: BTreeMap<String, String>, // query_id -> final_memory
pub event_count: u32,
pub chunks_seen: u32,
pub chunks_used: u32,
}
impl RebuildState {
/// Rebuild from event records (must be deterministic).
pub fn from_events(events: &[EventRecord]) -> Result<Self> {
let mut memories = BTreeMap::new();
let mut chunks_seen = 0;
let mut chunks_used = 0;
// Group events by query
let mut by_query: BTreeMap<String, Vec<&EventRecord>> = BTreeMap::new();
for event in events {
by_query.entry(event.query.clone()).or_insert_with(Vec::new).push(event);
}
// Replay events for each query
for (query_id, query_events) in by_query {
let memory = String::new();
let mut q_seen = 0;
let mut q_used = 0;
for event in query_events {
// Parse event_type (very simplified)
if event.event_type.contains("Memory") {
// Would parse the actual memory update from data
// For now: assume memory doesn't change without update
}
if event.event_type.contains("Evidence") {
q_used += 1;
}
if event.event_type.contains("Gate") {
q_seen += 1;
}
}
memories.insert(query_id, memory);
chunks_seen += q_seen;
chunks_used += q_used;
}
Ok(Self {
memories,
event_count: events.len() as u32,
chunks_seen,
chunks_used,
})
}
/// Serialize to JSONL (must match original byte-for-byte).
pub fn to_events(&self) -> Vec<EventRecord> {
// This is a placeholder - real rebuild would deserialize the exact events
// The key is that deserialization + re-serialization produces identical bytes
vec![]
}
}
#[cfg(test)]
mod tests {
use super::*;
use serde_json::json;
#[test]
fn test_rebuild_empty() {
let events = vec![];
let state = RebuildState::from_events(&events).unwrap();
assert_eq!(state.event_count, 0);
}
}