File size: 18,418 Bytes
aaf866d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
# Integration Guide β€” LightGBM Reranker into ResearchIT

> **For:** Whoever integrates the reranker into `app/recommend/reranker.py`  
> **Covers:** Steps 5-8 from the Phase 6 roadmap  
> **Prerequisites:** The production model is trained and in `production_model/reranker_v1.txt`

---

## Overview

You need to do 4 things:
1. **Expand `compute_features()` from 5 β†’ 37 features** (biggest change)
2. **Wire model loading + heuristic fallback** at startup
3. **Add `lightgbm` to `requirements.txt`** and model file to Docker image
4. **Integration testing**

---

## Step 5: Expand `compute_features()` to 37 Features

The current heuristic uses 5 features. The LightGBM model expects 37 features in a **specific order** defined in `production_model/feature_schema.json`.

### Feature Schema (order matters!)

```python
FEATURE_SCHEMA = [
    # Content/Retrieval (0-19)
    "qdrant_cosine_score",           # 0  - from Qdrant ANN search
    "candidate_position",            # 1  - rank in ANN results
    "candidate_citation_count",      # 2  - from Turso papers table
    "candidate_log_citations",       # 3  - log(citation_count + 1)
    "candidate_influential_citations",  # 4  - from Turso papers table
    "candidate_age_days",            # 5  - (now - update_date).days
    "candidate_recency_score",       # 6  - exp(-0.002 * age_days)
    "query_citation_count",          # 7  - user's profile paper citations (or 0)
    "query_age_days",                # 8  - user's profile paper age (or 0)
    "year_diff",                     # 9  - |query_year - candidate_year|
    "same_primary_category",         # 10 - 1 if same primary_topic
    "co_citation_count",             # 11 - shared citers (expensive; can be 0)
    "shared_author_count",           # 12 - shared authors between query & candidate
    "candidate_is_newer",            # 13 - 1 if candidate.year > query.year
    "query_log_citations",           # 14 - log(query_citation_count + 1)
    "citation_count_ratio",          # 15 - cand_citations / (query_citations + 1)
    "age_ratio",                     # 16 - cand_age / (query_age + 1)
    "candidate_citations_per_year",  # 17 - citations / max(age_years, 0.5)
    "query_num_references",          # 18 - 0 for now (needs citation graph in prod)
    "candidate_num_cited_by",        # 19 - 0 for now (needs citation graph in prod)

    # User Behavior (20-30) β€” from EWMA profiles, clusters, interactions
    "ewma_longterm_similarity",      # 20 - cos(candidate_emb, user.lt_profile)
    "ewma_shortterm_similarity",     # 21 - cos(candidate_emb, user.st_profile)
    "ewma_negative_similarity",      # 22 - cos(candidate_emb, user.neg_profile)
    "cluster_importance",            # 23 - cluster weight from Ward clustering
    "cluster_distance_to_medoid",    # 24 - cos(candidate_emb, cluster_medoid)
    "is_suppressed_category",        # 25 - 1 if suppressed category
    "onboarding_category_match",     # 26 - 1 if matches onboarding prefs
    "user_total_saves",              # 27 - total saves from interactions table
    "user_total_dismissals",         # 28 - total dismissals
    "user_days_since_last_save",     # 29 - days since last save
    "user_session_save_count",       # 30 - saves this session

    # Cross Features (31-36) β€” computed from above
    "cosine_x_recency",             # 31 - feat[0] * feat[6]
    "cosine_x_citations",           # 32 - feat[0] * feat[3]
    "category_x_recency",           # 33 - feat[10] * feat[6]
    "cosine_x_cocitation",          # 34 - feat[0] * log(feat[11] + 1)
    "position_inverse",             # 35 - 1 / (feat[1] + 1)
    "citations_x_recency",          # 36 - feat[3] * feat[6]
]
```

### Implementation Sketch

```python
import numpy as np
from datetime import datetime, timezone

def compute_features_v2(
    user_state: dict,        # EWMA profiles, cluster info, interaction counts
    candidate: dict,         # paper metadata from Turso
    qdrant_score: float,     # cosine score from ANN search
    candidate_position: int, # rank position (0-indexed)
    candidate_embedding: np.ndarray,  # 1024-dim BGE-M3 embedding
) -> np.ndarray:
    """
    Compute 37-feature vector for LightGBM reranker.
    
    Args:
        user_state: {
            "lt_profile": np.ndarray,       # long-term EWMA (1024-dim or None)
            "st_profile": np.ndarray,       # short-term EWMA (1024-dim or None)
            "neg_profile": np.ndarray,      # negative EWMA (1024-dim or None)
            "cluster_importance": float,     # from Ward clustering
            "cluster_medoid": np.ndarray,   # cluster medoid embedding (or None)
            "suppressed_categories": set,    # suppressed arXiv categories
            "onboarding_categories": set,    # onboarding selections
            "total_saves": int,
            "total_dismissals": int,
            "days_since_last_save": float,
            "session_save_count": int,
            "query_paper": dict | None,     # the "seed" paper if applicable
        }
        candidate: {
            "arxiv_id": str,
            "primary_topic": str,
            "update_date": str,              # "YYYY-MM-DD"
            "citation_count": int,
            "influential_citations": int,
            "authors": list[str],
        }
        qdrant_score: cosine similarity from ANN search
        candidate_position: rank in ANN results (0-indexed)
        candidate_embedding: paper's BGE-M3 embedding vector
        
    Returns:
        np.ndarray of shape (37,) β€” feature vector in schema order
    """
    features = np.zeros(37, dtype=np.float32)
    now = datetime.now(timezone.utc)
    
    # --- Content/Retrieval features (0-19) ---
    
    # 0: qdrant_cosine_score
    features[0] = qdrant_score
    
    # 1: candidate_position
    features[1] = float(candidate_position)
    
    # 2: candidate_citation_count
    cand_citations = candidate.get("citation_count", 0) or 0
    features[2] = float(cand_citations)
    
    # 3: candidate_log_citations
    features[3] = np.log(cand_citations + 1)
    
    # 4: candidate_influential_citations
    features[4] = float(candidate.get("influential_citations", 0) or 0)
    
    # 5: candidate_age_days
    try:
        pub_date = datetime.strptime(candidate.get("update_date", "")[:10], "%Y-%m-%d")
        pub_date = pub_date.replace(tzinfo=timezone.utc)
        cand_age = max(0, (now - pub_date).days)
    except (ValueError, TypeError):
        cand_age = 365  # default 1 year
    features[5] = float(cand_age)
    
    # 6: candidate_recency_score
    features[6] = np.exp(-0.002 * cand_age)
    
    # 7-9: Query paper features (from user's seed paper, or defaults)
    query_paper = user_state.get("query_paper") or {}
    query_citations = query_paper.get("citation_count", 0) or 0
    features[7] = float(query_citations)
    
    try:
        q_pub = datetime.strptime(query_paper.get("update_date", "")[:10], "%Y-%m-%d")
        q_pub = q_pub.replace(tzinfo=timezone.utc)
        query_age = max(0, (now - q_pub).days)
    except (ValueError, TypeError):
        query_age = 0
    features[8] = float(query_age)
    
    cand_year = _parse_year(candidate.get("update_date", ""))
    query_year = _parse_year(query_paper.get("update_date", "")) if query_paper else cand_year
    features[9] = abs(query_year - cand_year)
    
    # 10: same_primary_category
    q_cat = query_paper.get("primary_topic", "") if query_paper else ""
    c_cat = candidate.get("primary_topic", "")
    features[10] = 1.0 if (q_cat and c_cat and q_cat == c_cat) else 0.0
    
    # 11: co_citation_count (0 unless you have citation graph loaded)
    features[11] = 0.0  # TODO: populate if citation graph is loaded
    
    # 12: shared_author_count
    if query_paper and query_paper.get("authors"):
        q_authors = {a.lower().strip() for a in query_paper["authors"] if a}
        c_authors = {a.lower().strip() for a in (candidate.get("authors") or []) if a}
        features[12] = float(len(q_authors & c_authors))
    
    # 13: candidate_is_newer
    features[13] = 1.0 if cand_year > query_year else 0.0
    
    # 14: query_log_citations
    features[14] = np.log(query_citations + 1)
    
    # 15: citation_count_ratio
    features[15] = cand_citations / (query_citations + 1)
    
    # 16: age_ratio
    features[16] = cand_age / (query_age + 1) if query_age > 0 else 0.0
    
    # 17: candidate_citations_per_year
    cand_age_years = max(cand_age / 365.0, 0.5)
    features[17] = cand_citations / cand_age_years
    
    # 18-19: Graph features (0 unless citation graph loaded in prod)
    features[18] = 0.0  # query_num_references
    features[19] = 0.0  # candidate_num_cited_by
    
    # --- User Behavior features (20-30) ---
    
    # 20: ewma_longterm_similarity
    lt_prof = user_state.get("lt_profile")
    if lt_prof is not None and candidate_embedding is not None:
        features[20] = _cosine_sim(candidate_embedding, lt_prof)
    
    # 21: ewma_shortterm_similarity
    st_prof = user_state.get("st_profile")
    if st_prof is not None and candidate_embedding is not None:
        features[21] = _cosine_sim(candidate_embedding, st_prof)
    
    # 22: ewma_negative_similarity
    neg_prof = user_state.get("neg_profile")
    if neg_prof is not None and candidate_embedding is not None:
        features[22] = _cosine_sim(candidate_embedding, neg_prof)
    
    # 23: cluster_importance
    features[23] = float(user_state.get("cluster_importance", 0.0))
    
    # 24: cluster_distance_to_medoid
    medoid = user_state.get("cluster_medoid")
    if medoid is not None and candidate_embedding is not None:
        features[24] = _cosine_sim(candidate_embedding, medoid)
    
    # 25: is_suppressed_category
    suppressed = user_state.get("suppressed_categories", set())
    features[25] = 1.0 if c_cat in suppressed else 0.0
    
    # 26: onboarding_category_match
    onboarding = user_state.get("onboarding_categories", set())
    features[26] = 1.0 if c_cat in onboarding else 0.0
    
    # 27-30: Interaction counts
    features[27] = float(user_state.get("total_saves", 0))
    features[28] = float(user_state.get("total_dismissals", 0))
    features[29] = float(user_state.get("days_since_last_save", 0.0))
    features[30] = float(user_state.get("session_save_count", 0))
    
    # --- Cross Features (31-36) ---
    
    features[31] = features[0] * features[6]   # cosine Γ— recency
    features[32] = features[0] * features[3]   # cosine Γ— log_citations
    features[33] = features[10] * features[6]  # category Γ— recency
    features[34] = features[0] * np.log(features[11] + 1)  # cosine Γ— log_cocitation
    features[35] = 1.0 / (features[1] + 1)     # position_inverse
    features[36] = features[3] * features[6]   # log_citations Γ— recency
    
    return features


def _cosine_sim(a: np.ndarray, b: np.ndarray) -> float:
    """Cosine similarity between two vectors."""
    dot = np.dot(a, b)
    norm_a = np.linalg.norm(a)
    norm_b = np.linalg.norm(b)
    if norm_a == 0 or norm_b == 0:
        return 0.0
    return float(dot / (norm_a * norm_b))


def _parse_year(date_str: str) -> int:
    try:
        return int(date_str[:4])
    except (ValueError, TypeError, IndexError):
        return 2020
```

### Vectorized Version (for batch scoring)

For production use, compute features for ALL candidates at once:

```python
def compute_features_batch(
    user_state: dict,
    candidates: list[dict],
    qdrant_scores: list[float],
    candidate_embeddings: np.ndarray,  # (N, 1024)
) -> np.ndarray:
    """
    Compute features for all candidates at once.
    Returns (N, 37) feature matrix.
    """
    N = len(candidates)
    features = np.zeros((N, 37), dtype=np.float32)
    
    for i, (cand, score) in enumerate(zip(candidates, qdrant_scores)):
        features[i] = compute_features_v2(
            user_state=user_state,
            candidate=cand,
            qdrant_score=score,
            candidate_position=i,
            candidate_embedding=candidate_embeddings[i] if candidate_embeddings is not None else None,
        )
    
    return features
```

> **Performance note:** The bottleneck is NOT feature computation or LightGBM prediction (0.4ms). It's fetching candidate metadata from Turso. Batch your Turso queries.

---

## Step 6: Wire Model Loading + Heuristic Fallback

In `app/recommend/reranker.py`:

```python
import os
import lightgbm as lgb
import numpy as np

# ── Model Loading ────────────────────────────────────────────────────────────

_lgb_model = None
_model_path = os.environ.get("RERANKER_MODEL_PATH", "production_model/reranker_v1.txt")

try:
    _lgb_model = lgb.Booster(model_file=_model_path)
    print(f"[reranker] LightGBM model loaded from {_model_path}")
    print(f"[reranker]   num_features: {_lgb_model.num_feature()}")
    print(f"[reranker]   num_trees: {_lgb_model.num_trees()}")
except FileNotFoundError:
    print(f"[reranker] Model file not found: {_model_path} β€” using heuristic")
except Exception as e:
    print(f"[reranker] Model load failed: {e} β€” using heuristic")


# ── Main Reranking Function ──────────────────────────────────────────────────

def rerank_candidates(
    user_state: dict,
    candidates: list[dict],
    qdrant_scores: list[float],
    candidate_embeddings: np.ndarray | None = None,
) -> list[dict]:
    """
    Rerank candidates using LightGBM (or heuristic fallback).
    
    Returns candidates sorted by score (best first).
    """
    if not candidates:
        return []
    
    if _lgb_model is not None:
        # LightGBM path
        features = compute_features_batch(user_state, candidates, qdrant_scores, candidate_embeddings)
        scores = _lgb_model.predict(features)
    else:
        # Heuristic fallback (always works, no model needed)
        scores = np.array([
            heuristic_score(user_state, cand, score)
            for cand, score in zip(candidates, qdrant_scores)
        ])
    
    # Sort by score descending
    order = np.argsort(-scores)
    return [candidates[i] for i in order]
```

### Key Design Decisions

1. **The heuristic fallback is PERMANENT.** Don't remove it. It's your safety net if:
   - The model file is missing (fresh deploy)
   - LightGBM import fails (dependency issue)
   - The model produces garbage (bad retrain)

2. **Model path is configurable** via `RERANKER_MODEL_PATH` env var. This lets you A/B test different models without code changes.

3. **No model versioning yet.** For v1, just replace the file. When you have v2, add version tracking.

---

## Step 7: Update `requirements.txt`

Add to your `requirements.txt`:
```
lightgbm>=4.0,<5.0
```

And in your `Dockerfile`, ensure the model file is copied:
```dockerfile
COPY production_model/reranker_v1.txt /app/production_model/reranker_v1.txt
```

Or download from this repo at startup:
```python
# In app startup
from huggingface_hub import hf_hub_download

model_path = hf_hub_download(
    repo_id="siddhm11/researchit-reranker-phase6",
    filename="production_model/reranker_v1.txt",
)
```

---

## Step 8: Integration Testing

### Smoke Test
```python
import lightgbm as lgb
import numpy as np

# Load model
model = lgb.Booster(model_file="production_model/reranker_v1.txt")
assert model.num_feature() == 37

# Predict on dummy input
dummy = np.zeros((5, 37), dtype=np.float32)
scores = model.predict(dummy)
assert scores.shape == (5,)
assert not np.any(np.isnan(scores))
print("βœ… Smoke test passed")
```

### End-to-End Test
```python
# Verify the full pipeline: ANN β†’ feature computation β†’ LightGBM β†’ ranked output
def test_e2e():
    # 1. Simulate a user with EWMA profiles
    user_state = {
        "lt_profile": np.random.randn(1024).astype(np.float32),
        "st_profile": np.random.randn(1024).astype(np.float32),
        "neg_profile": np.random.randn(1024).astype(np.float32),
        "cluster_importance": 0.8,
        "cluster_medoid": np.random.randn(1024).astype(np.float32),
        "suppressed_categories": {"cs.CR"},
        "onboarding_categories": {"cs.CL", "cs.LG"},
        "total_saves": 42,
        "total_dismissals": 10,
        "days_since_last_save": 0.5,
        "session_save_count": 3,
        "query_paper": None,
    }
    
    # 2. Simulate candidates from Qdrant
    candidates = [
        {"arxiv_id": f"2024.{i:05d}", "primary_topic": "cs.CL",
         "update_date": "2024-01-15", "citation_count": i*10,
         "influential_citations": i, "authors": ["Alice", "Bob"]}
        for i in range(50)
    ]
    qdrant_scores = [0.9 - i*0.01 for i in range(50)]
    candidate_embeddings = np.random.randn(50, 1024).astype(np.float32)
    
    # 3. Rerank
    ranked = rerank_candidates(user_state, candidates, qdrant_scores, candidate_embeddings)
    
    assert len(ranked) == 50
    # The order should differ from the ANN order (LightGBM reranks)
    original_ids = [c["arxiv_id"] for c in candidates]
    reranked_ids = [c["arxiv_id"] for c in ranked]
    assert original_ids != reranked_ids, "LightGBM should change the order"
    print("βœ… E2E test passed")
```

### Latency Test
```python
import time

features = np.random.randn(100, 37).astype(np.float32)

# Warmup
for _ in range(100):
    model.predict(features)

# Benchmark
t0 = time.time()
for _ in range(1000):
    model.predict(features)
elapsed = (time.time() - t0) / 1000 * 1000  # ms per call

assert elapsed < 1.0, f"Too slow: {elapsed:.3f}ms (target: <1ms)"
print(f"βœ… Latency: {elapsed:.3f}ms per 100 candidates")
```

---

## Notes for Future Retraining

When you have 500+ real user interactions:

1. Export interactions from Turso:
   ```sql
   SELECT user_id, arxiv_id, action, created_at FROM interactions
   ```

2. Generate new training triples with **real labels**:
   - `action = 'save'` β†’ label 2
   - `action = 'click'` β†’ label 1
   - `action = 'dismiss'` β†’ label 0

3. The 37-feature schema is **stable** β€” features 20-30 will now be populated with real EWMA profiles, cluster data, and interaction counts.

4. Retrain with the same `03_train_lightgbm.py` script on the new data.

5. The user behavior features (20-30) should gain significant importance in the new model.