memd 1.9.0

Local memory CLI for AI agents with hybrid retrieval
Documentation
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
//! Served-versus-shadow counterfactual evaluation for outcome-v1 ranking.

use std::collections::HashSet;
use std::path::{Path, PathBuf};
use std::time::{SystemTime, UNIX_EPOCH};

use serde::{Deserialize, Serialize};
use serde_json::{json, Value};
use sha2::{Digest, Sha256};

use super::paths::absolutize_project_dir;
use super::unwrap_content_payload;
use crate::error::{MemdError, Result};
use crate::mcp::handlers::{handle_memory_search, SearchParams};
use crate::store::{RankingPolicyMode, RetrievalEpisodeId, RetrievalEpisodeItem, Store};
use crate::types::TenantId;

#[derive(Debug, Clone)]
pub(super) struct EvalOutcomeRankingOptions {
    pub(super) tenant_id: String,
    pub(super) project_id: Option<String>,
    pub(super) project_dir: PathBuf,
    pub(super) queries_path: PathBuf,
    pub(super) k: usize,
    pub(super) report_json: PathBuf,
}

#[derive(Debug, Clone, Deserialize)]
struct OutcomeBenchQuery {
    id: String,
    query: String,
    #[serde(default)]
    relevant_chunk_ids: Vec<String>,
    #[serde(default)]
    harmful_chunk_ids: Vec<String>,
}

#[derive(Debug, Clone, Serialize)]
struct PolicyMetrics {
    recall_at_k: f64,
    reciprocal_rank: f64,
    harmful_at_k: usize,
}

#[derive(Debug, Clone, Serialize)]
struct OutcomeCounterfactualRow {
    query_id: String,
    query_hash: String,
    retrieval_episode_id: String,
    served_top_k: Vec<String>,
    shadow_top_k: Vec<String>,
    order_changed: bool,
    served: PolicyMetrics,
    shadow: PolicyMetrics,
}

#[derive(Debug, Clone, Serialize)]
struct AggregatePolicyMetrics {
    mean_recall_at_k: f64,
    mean_reciprocal_rank: f64,
    mean_harmful_at_k: f64,
}

#[derive(Debug, Clone, Serialize)]
struct OutcomeCounterfactualSummary {
    query_count: usize,
    changed_query_count: usize,
    served: AggregatePolicyMetrics,
    shadow: AggregatePolicyMetrics,
    shadow_minus_served_recall_at_k: f64,
    shadow_minus_served_mrr: f64,
    shadow_minus_served_harmful_at_k: f64,
}

#[derive(Debug, Clone, Serialize)]
struct OutcomeCounterfactualReport {
    schema_version: &'static str,
    generated_unix_ms: u128,
    tenant_id: String,
    project_id: Option<String>,
    policy_version: &'static str,
    policy_mode: RankingPolicyMode,
    k: usize,
    queries_path: String,
    queries_sha256: String,
    summary: OutcomeCounterfactualSummary,
    rows: Vec<OutcomeCounterfactualRow>,
}

pub(super) async fn run_eval_outcome_ranking<S: Store>(
    store: &S,
    options: EvalOutcomeRankingOptions,
) -> Result<Value> {
    let project_dir = absolutize_project_dir(&options.project_dir)?;
    let queries_path = resolve_from(&project_dir, &options.queries_path);
    let report_json = resolve_from(&project_dir, &options.report_json);
    let report_markdown = report_json.with_extension("md");
    refuse_overwrite(&report_json)?;
    refuse_overwrite(&report_markdown)?;
    let tenant_id = TenantId::new(&options.tenant_id)?;
    let k = options.k.clamp(1, 50);
    let (queries, queries_sha256) = load_queries(&queries_path)?;
    if queries.is_empty() {
        return Err(MemdError::ValidationError(format!(
            "no outcome-ranking queries found in {}",
            queries_path.display()
        )));
    }

    let mut rows = Vec::with_capacity(queries.len());
    for query in queries {
        let response = handle_memory_search(
            store,
            SearchParams {
                tenant_id: tenant_id.to_string(),
                project_id: options.project_id.clone(),
                query: query.query.clone(),
                k,
                dedupe_by_source: true,
                include_text: Some(false),
                ranking_policy: Some(RankingPolicyMode::Shadow),
                candidate_multiplier: Some(4),
                task_id: Some(format!("eval-outcome-ranking:{}", query.id)),
                suppress_usage_event: true,
                ..Default::default()
            },
        )
        .await
        .map_err(|error| MemdError::ProtocolError(error.to_string()))?;
        let payload = unwrap_content_payload(response)?;
        let episode_id = payload
            .get("retrieval_episode_id")
            .and_then(Value::as_str)
            .ok_or_else(|| {
                MemdError::ProtocolError(
                    "outcome-ranking evaluation search returned no retrieval episode".to_string(),
                )
            })?;
        let parsed_episode_id = RetrievalEpisodeId::parse(episode_id)?;
        let (_, items) = store
            .get_retrieval_episode(&tenant_id, &parsed_episode_id)
            .await?
            .ok_or_else(|| {
                MemdError::StorageError(format!(
                    "retrieval episode {episode_id} disappeared during evaluation"
                ))
            })?;
        let served_top_k = ranked_chunk_ids(&items, k, false);
        let shadow_top_k = ranked_chunk_ids(&items, k, true);
        let served = policy_metrics(
            &served_top_k,
            &query.relevant_chunk_ids,
            &query.harmful_chunk_ids,
        );
        let shadow = policy_metrics(
            &shadow_top_k,
            &query.relevant_chunk_ids,
            &query.harmful_chunk_ids,
        );
        rows.push(OutcomeCounterfactualRow {
            query_id: query.id,
            query_hash: crate::store::stable_query_hash(&query.query),
            retrieval_episode_id: episode_id.to_string(),
            order_changed: served_top_k != shadow_top_k,
            served_top_k,
            shadow_top_k,
            served,
            shadow,
        });
    }

    let summary = summarize(&rows);
    let report = OutcomeCounterfactualReport {
        schema_version: "memd.outcome_counterfactual.v1",
        generated_unix_ms: now_ms(),
        tenant_id: tenant_id.to_string(),
        project_id: options.project_id,
        policy_version: crate::store::OUTCOME_POLICY_VERSION,
        policy_mode: RankingPolicyMode::Shadow,
        k,
        queries_path: queries_path.display().to_string(),
        queries_sha256,
        summary,
        rows,
    };
    if let Some(parent) = report_json.parent() {
        std::fs::create_dir_all(parent)?;
    }
    std::fs::write(&report_json, serde_json::to_string_pretty(&report)?)?;
    std::fs::write(&report_markdown, render_markdown(&report))?;

    Ok(json!({
        "schema_version": report.schema_version,
        "queries": report.summary.query_count,
        "changed_queries": report.summary.changed_query_count,
        "shadow_minus_served_recall_at_k": report.summary.shadow_minus_served_recall_at_k,
        "shadow_minus_served_mrr": report.summary.shadow_minus_served_mrr,
        "shadow_minus_served_harmful_at_k": report.summary.shadow_minus_served_harmful_at_k,
        "report_json": report_json,
        "report_markdown": report_markdown,
    }))
}

fn resolve_from(project_dir: &Path, path: &Path) -> PathBuf {
    if path.is_absolute() {
        path.to_path_buf()
    } else {
        project_dir.join(path)
    }
}

fn refuse_overwrite(path: &Path) -> Result<()> {
    if path.exists() {
        return Err(MemdError::ValidationError(format!(
            "refusing to overwrite outcome-ranking artifact {}",
            path.display()
        )));
    }
    Ok(())
}

fn load_queries(path: &Path) -> Result<(Vec<OutcomeBenchQuery>, String)> {
    let bytes = std::fs::read(path).map_err(|error| {
        MemdError::ValidationError(format!(
            "cannot read outcome-ranking queries {}: {error}",
            path.display()
        ))
    })?;
    let sha256 = format!("{:x}", Sha256::digest(&bytes));
    let text = String::from_utf8(bytes).map_err(|error| {
        MemdError::ValidationError(format!(
            "outcome-ranking queries {} are not UTF-8: {error}",
            path.display()
        ))
    })?;
    let mut queries = Vec::new();
    let mut ids = HashSet::new();
    for (index, line) in text.lines().enumerate() {
        let line = line.trim();
        if line.is_empty() || line.starts_with('#') {
            continue;
        }
        let query: OutcomeBenchQuery = serde_json::from_str(line).map_err(|error| {
            MemdError::ValidationError(format!(
                "invalid outcome-ranking JSON on line {} of {}: {error}",
                index + 1,
                path.display()
            ))
        })?;
        validate_query(&query, &mut ids, index + 1)?;
        queries.push(query);
    }
    Ok((queries, sha256))
}

fn validate_query(query: &OutcomeBenchQuery, ids: &mut HashSet<String>, line: usize) -> Result<()> {
    if query.id.trim().is_empty() || query.query.trim().is_empty() {
        return Err(MemdError::ValidationError(format!(
            "outcome-ranking line {line} requires non-empty id and query"
        )));
    }
    if !ids.insert(query.id.clone()) {
        return Err(MemdError::ValidationError(format!(
            "duplicate outcome-ranking query id {}",
            query.id
        )));
    }
    let relevant = query.relevant_chunk_ids.iter().collect::<HashSet<_>>();
    let harmful = query.harmful_chunk_ids.iter().collect::<HashSet<_>>();
    if relevant.len() != query.relevant_chunk_ids.len()
        || harmful.len() != query.harmful_chunk_ids.len()
        || relevant.iter().any(|id| harmful.contains(id))
    {
        return Err(MemdError::ValidationError(format!(
            "outcome-ranking query {} has duplicate or overlapping judgments",
            query.id
        )));
    }
    Ok(())
}

fn ranked_chunk_ids(items: &[RetrievalEpisodeItem], k: usize, shadow: bool) -> Vec<String> {
    let mut ranked = items
        .iter()
        .filter_map(|item| {
            let rank = if shadow {
                item.shadow_rank
            } else {
                item.served_rank
            }?;
            Some((rank, item))
        })
        .collect::<Vec<_>>();
    ranked.sort_by_key(|(rank, _)| *rank);
    let mut source_groups = HashSet::new();
    let mut chunks = Vec::new();
    for (_, item) in ranked {
        if let Some(group) = item.source_dedup_group.as_deref() {
            if !source_groups.insert(group) {
                continue;
            }
        }
        chunks.push(item.chunk_id.to_string());
        if chunks.len() >= k {
            break;
        }
    }
    chunks
}

fn policy_metrics(order: &[String], relevant: &[String], harmful: &[String]) -> PolicyMetrics {
    let relevant = relevant.iter().collect::<HashSet<_>>();
    let harmful = harmful.iter().collect::<HashSet<_>>();
    let relevant_hits = order.iter().filter(|id| relevant.contains(id)).count();
    let recall_at_k = if relevant.is_empty() {
        0.0
    } else {
        relevant_hits as f64 / relevant.len() as f64
    };
    let reciprocal_rank = order
        .iter()
        .position(|id| relevant.contains(id))
        .map(|rank| 1.0 / (rank + 1) as f64)
        .unwrap_or(0.0);
    let harmful_at_k = order.iter().filter(|id| harmful.contains(id)).count();
    PolicyMetrics {
        recall_at_k,
        reciprocal_rank,
        harmful_at_k,
    }
}

fn summarize(rows: &[OutcomeCounterfactualRow]) -> OutcomeCounterfactualSummary {
    let query_count = rows.len();
    let mean = |values: Vec<f64>| values.iter().sum::<f64>() / query_count as f64;
    let aggregate = |shadow: bool| AggregatePolicyMetrics {
        mean_recall_at_k: mean(
            rows.iter()
                .map(|row| {
                    if shadow {
                        row.shadow.recall_at_k
                    } else {
                        row.served.recall_at_k
                    }
                })
                .collect(),
        ),
        mean_reciprocal_rank: mean(
            rows.iter()
                .map(|row| {
                    if shadow {
                        row.shadow.reciprocal_rank
                    } else {
                        row.served.reciprocal_rank
                    }
                })
                .collect(),
        ),
        mean_harmful_at_k: mean(
            rows.iter()
                .map(|row| {
                    if shadow {
                        row.shadow.harmful_at_k as f64
                    } else {
                        row.served.harmful_at_k as f64
                    }
                })
                .collect(),
        ),
    };
    let served = aggregate(false);
    let shadow = aggregate(true);
    OutcomeCounterfactualSummary {
        query_count,
        changed_query_count: rows.iter().filter(|row| row.order_changed).count(),
        shadow_minus_served_recall_at_k: shadow.mean_recall_at_k - served.mean_recall_at_k,
        shadow_minus_served_mrr: shadow.mean_reciprocal_rank - served.mean_reciprocal_rank,
        shadow_minus_served_harmful_at_k: shadow.mean_harmful_at_k - served.mean_harmful_at_k,
        served,
        shadow,
    }
}

fn render_markdown(report: &OutcomeCounterfactualReport) -> String {
    let mut output = String::new();
    output.push_str("# Outcome-ranking counterfactual\n\n");
    output.push_str(&format!(
        "- policy: `{}` (`shadow`)\n",
        report.policy_version
    ));
    output.push_str(&format!("- queries: `{}`\n", report.summary.query_count));
    output.push_str(&format!("- k: `{}`\n", report.k));
    output.push_str(&format!(
        "- changed queries: `{}`\n",
        report.summary.changed_query_count
    ));
    output.push_str(&format!(
        "- shadow - served recall@k: `{:.4}`\n",
        report.summary.shadow_minus_served_recall_at_k
    ));
    output.push_str(&format!(
        "- shadow - served MRR: `{:.4}`\n",
        report.summary.shadow_minus_served_mrr
    ));
    output.push_str(&format!(
        "- shadow - served harmful@k: `{:.4}`\n\n",
        report.summary.shadow_minus_served_harmful_at_k
    ));
    output.push_str("| query_id | changed | served recall | shadow recall | served harmful | shadow harmful |\n");
    output.push_str("|---|---:|---:|---:|---:|---:|\n");
    for row in &report.rows {
        output.push_str(&format!(
            "| {} | {} | {:.3} | {:.3} | {} | {} |\n",
            row.query_id.replace('|', "\\|"),
            row.order_changed,
            row.served.recall_at_k,
            row.shadow.recall_at_k,
            row.served.harmful_at_k,
            row.shadow.harmful_at_k,
        ));
    }
    output
}

fn now_ms() -> u128 {
    SystemTime::now()
        .duration_since(UNIX_EPOCH)
        .map(|duration| duration.as_millis())
        .unwrap_or(0)
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn metrics_distinguish_relevant_and_harmful_ranks() {
        let metrics = policy_metrics(
            &["harm".to_string(), "ok".to_string()],
            &["ok".to_string()],
            &["harm".to_string()],
        );
        assert_eq!(metrics.recall_at_k, 1.0);
        assert_eq!(metrics.reciprocal_rank, 0.5);
        assert_eq!(metrics.harmful_at_k, 1);
    }

    #[test]
    fn shadow_order_applies_source_dedup() {
        let episode_id = RetrievalEpisodeId::new();
        let tenant_id = TenantId::new("t").unwrap();
        let make = |id: &str, rank: usize, group: Option<&str>| RetrievalEpisodeItem {
            episode_id: episode_id.clone(),
            chunk_id: crate::types::ChunkId::parse(id).unwrap(),
            origin_tenant_id: tenant_id.clone(),
            origin_project_id: None,
            original_rank: rank,
            original_score: 1.0,
            lane_scores_json: "{}".to_string(),
            outcome_adjustment: 0.0,
            served_rank: None,
            shadow_rank: Some(rank),
            rendered: false,
            source_dedup_group: group.map(str::to_string),
        };
        let items = vec![
            make("01900000-0000-7000-8000-000000000001", 0, Some("same")),
            make("01900000-0000-7000-8000-000000000002", 1, Some("same")),
            make("01900000-0000-7000-8000-000000000003", 2, None),
        ];
        assert_eq!(ranked_chunk_ids(&items, 2, true).len(), 2);
        assert_eq!(
            ranked_chunk_ids(&items, 2, true)[1],
            "01900000-0000-7000-8000-000000000003"
        );
    }
}