my prefect server setup
prefect-metrics.waow.tech
python
orchestration
fix weave: LLM-first tag consolidation instead of embedding pre-filter
give the LLM the full tag inventory with context so it can propose merges
and relationships directly, rather than only seeing pre-filtered pairs that
passed an embedding similarity threshold. the previous approach found zero
merges because short tag strings don't cluster well by cosine similarity.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
+75
-79
+75
-79
flows/weave.py
+75
-79
flows/weave.py
···
185
185
tag_embeddings: dict[str, list[float]],
186
186
api_key: str,
187
187
) -> list[dict[str, Any]]:
188
-
"""Cluster tags by embedding similarity, LLM confirms merges."""
189
-
tags = list(tag_embeddings.keys())
188
+
"""Give the LLM the full tag inventory and let it propose consolidations."""
189
+
# format every tag with usage context so the LLM can make informed decisions
190
+
tag_lines = []
191
+
for tag in sorted(tag_info.keys()):
192
+
info = tag_info[tag]
193
+
users = info.get("users", [])
194
+
count = info.get("count", 0)
195
+
episodic = info.get("episodic_count", 0)
196
+
sample = (info.get("samples") or [""])[0][:120]
197
+
tag_lines.append(
198
+
f" {tag} (obs={count}, episodic={episodic}, users={len(users)})"
199
+
f"\n sample: {sample}"
200
+
)
190
201
191
-
# find high-similarity pairs (>= 0.85) as merge candidates
192
-
candidates: list[tuple[str, str, float]] = []
193
-
for i, t1 in enumerate(tags):
194
-
for t2 in tags[i + 1 :]:
195
-
sim = cosine_similarity(tag_embeddings[t1], tag_embeddings[t2])
196
-
if sim >= 0.85:
197
-
candidates.append((t1, t2, sim))
198
-
199
-
if not candidates:
200
-
return []
201
-
202
-
candidates.sort(key=lambda x: -x[2])
203
-
candidates_text = "\n".join(
204
-
f"- \"{t1}\" <-> \"{t2}\" (similarity: {sim:.3f})\n"
205
-
f" {t1} context: {(tag_info.get(t1) or {}).get('samples', [''])[0][:100]}\n"
206
-
f" {t2} context: {(tag_info.get(t2) or {}).get('samples', [''])[0][:100]}"
207
-
for t1, t2, sim in candidates[:30] # cap at 30 pairs
208
-
)
202
+
inventory = "\n".join(tag_lines)
209
203
210
204
model = AnthropicModel(
211
205
"claude-haiku-4-5", provider=AnthropicProvider(api_key=api_key)
···
213
207
agent = Agent(
214
208
model,
215
209
system_prompt=(
216
-
"you review tag merge candidates for a memory graph. for each pair, decide:\n"
217
-
"- MERGE: same concept — pick the canonical form\n"
218
-
"- RELATE: distinct but related — don't merge\n"
219
-
"- SKIP: not meaningfully related despite embedding similarity\n\n"
220
-
"prefer lowercase, hyphenated canonical forms (e.g. 'ai-systems' not 'AI_systems').\n"
221
-
"group transitive merges (if a merges with b and b merges with c, "
222
-
"produce one merge with canonical + all aliases).\n"
223
-
"put RELATE pairs in the 'related' field of the merge they're closest to, "
224
-
"or omit if they don't belong to any merge group."
210
+
"you are consolidating the tag vocabulary for phi's memory graph.\n\n"
211
+
"you will receive the full inventory of tags with usage counts and sample "
212
+
"observations. your job:\n\n"
213
+
"1. MERGE tags that are the same concept with different surface forms.\n"
214
+
" examples: 'attestation' / 'self-attestation' → canonical: 'attestation'\n"
215
+
" 'ai_systems' / 'bot' / 'system-improvement' → canonical: 'ai-systems'\n\n"
216
+
"2. mark tags that are RELATED but distinct — these should link, not merge.\n"
217
+
" example: 'epistemology' / 'social-epistemology' → related, not merged\n\n"
218
+
"rules:\n"
219
+
"- prefer lowercase, hyphenated canonical forms\n"
220
+
"- group transitive merges into one entry\n"
221
+
"- put related (but not merged) tags in the 'related' field\n"
222
+
"- only merge when you're confident they're the same concept\n"
223
+
"- it's fine to return zero merges if the tags are already clean\n"
224
+
"- look for underscored vs hyphenated variants, singular/plural, "
225
+
"abbreviations, and overlapping concepts"
225
226
),
226
227
output_type=MergeProposal,
227
228
name="tag-merger",
228
229
)
229
230
230
-
result = await agent.run(f"merge candidates:\n{candidates_text}")
231
-
# serialize to dicts for prefect result persistence
231
+
result = await agent.run(f"full tag inventory ({len(tag_info)} tags):\n{inventory}")
232
232
return [m.model_dump() for m in result.output.merges]
233
233
234
234
···
353
353
merged_aliases: set[str],
354
354
api_key: str,
355
355
) -> list[dict[str, Any]]:
356
-
"""Score and LLM-confirm relationships between non-merged tags."""
357
-
tags = [t for t in tag_embeddings if t not in merged_aliases]
358
-
359
-
# score tag pairs by combined signal
360
-
scored: list[tuple[str, str, float, str]] = [] # (a, b, score, reason)
361
-
for i, t1 in enumerate(tags):
362
-
for t2 in tags[i + 1 :]:
363
-
sim = cosine_similarity(tag_embeddings[t1], tag_embeddings[t2])
364
-
if sim < 0.4:
365
-
continue
366
-
367
-
# co-occurrence score
368
-
pair_key = "|".join(sorted([t1, t2]))
369
-
cooccur = cooccurrences.get(pair_key, 0)
370
-
371
-
# shared users score
372
-
shared_users = sum(
373
-
1
374
-
for tags_list in user_tag_sets.values()
375
-
if t1 in tags_list and t2 in tags_list
376
-
)
377
-
378
-
# combine signals
379
-
score = sim * 0.5
380
-
if cooccur > 0:
381
-
score += min(cooccur / 5, 0.3) # cap at 0.3
382
-
if shared_users > 0:
383
-
score += min(shared_users / 3, 0.2) # cap at 0.2
356
+
"""Give the LLM the full tag list with co-occurrence context to find relationships."""
357
+
tags = [t for t in sorted(tag_info.keys()) if t not in merged_aliases]
384
358
385
-
if score >= 0.5:
386
-
reason = f"sim={sim:.2f}, cooccur={cooccur}, shared_users={shared_users}"
387
-
scored.append((t1, t2, score, reason))
359
+
# precompute co-occurrence hints to give the LLM as context
360
+
cooccur_hints: dict[str, list[str]] = defaultdict(list)
361
+
for pair_key, count in cooccurrences.items():
362
+
t1, t2 = pair_key.split("|", 1)
363
+
if t1 in merged_aliases or t2 in merged_aliases:
364
+
continue
365
+
if count >= 2:
366
+
cooccur_hints[t1].append(f"{t2} ({count}x)")
367
+
cooccur_hints[t2].append(f"{t1} ({count}x)")
388
368
389
-
if not scored:
390
-
return []
369
+
# format tag inventory with context
370
+
tag_lines = []
371
+
for tag in tags:
372
+
info = tag_info.get(tag, {})
373
+
count = info.get("count", 0)
374
+
episodic = info.get("episodic_count", 0)
375
+
n_users = len(info.get("users", []))
376
+
sample = (info.get("samples") or [""])[0][:120]
377
+
cooccur_str = ""
378
+
if tag in cooccur_hints:
379
+
cooccur_str = f"\n co-occurs with: {', '.join(cooccur_hints[tag][:5])}"
380
+
tag_lines.append(
381
+
f" {tag} (obs={count}, episodic={episodic}, users={n_users})"
382
+
f"\n sample: {sample}{cooccur_str}"
383
+
)
391
384
392
-
scored.sort(key=lambda x: -x[2])
393
-
candidates_text = "\n".join(
394
-
f"- \"{t1}\" <-> \"{t2}\" (score: {score:.2f}, {reason})\n"
395
-
f" {t1}: {(tag_info.get(t1) or {}).get('samples', [''])[0][:100]}\n"
396
-
f" {t2}: {(tag_info.get(t2) or {}).get('samples', [''])[0][:100]}"
397
-
for t1, t2, score, reason in scored[:30]
398
-
)
385
+
inventory = "\n".join(tag_lines)
399
386
400
387
model = AnthropicModel(
401
388
"claude-haiku-4-5", provider=AnthropicProvider(api_key=api_key)
···
403
390
agent = Agent(
404
391
model,
405
392
system_prompt=(
406
-
"you review candidate tag relationships for a memory graph belonging to phi, "
407
-
"a bluesky bot that remembers conversations.\n\n"
408
-
"for each pair, decide if there's a genuine conceptual relationship:\n"
393
+
"you are mapping relationships between tags in phi's memory graph.\n"
394
+
"phi is a bluesky bot that remembers conversations and builds knowledge.\n\n"
395
+
"you will receive the full tag inventory with usage counts, sample observations, "
396
+
"and co-occurrence data. your job: identify genuine conceptual relationships.\n\n"
397
+
"relationship types:\n"
409
398
"- RELATED: broadly connected concepts\n"
410
399
"- SUBTOPIC: one is a narrower form of the other\n"
411
-
"- OVERLAPPING: partially shared meaning\n"
412
-
"- SKIP: co-occurrence is coincidental, not conceptual\n\n"
413
-
"assign confidence 0.0-1.0. be honest — surface co-occurrence ≠ real relationship.\n"
414
-
"provide brief evidence for each accepted relationship."
400
+
"- OVERLAPPING: partially shared meaning but distinct\n\n"
401
+
"rules:\n"
402
+
"- assign confidence 0.0-1.0 based on how strong the connection is\n"
403
+
"- co-occurrence is a signal but not proof — two tags appearing together "
404
+
"might be coincidental\n"
405
+
"- look for thematic clusters, not just pairs\n"
406
+
"- provide brief evidence for each relationship\n"
407
+
"- be selective — only include relationships you're genuinely confident about\n"
408
+
"- skip trivially obvious connections (like a tag co-occurring with itself)"
415
409
),
416
410
output_type=RelationshipProposal,
417
411
name="tag-relator",
418
412
)
419
413
420
-
result = await agent.run(f"relationship candidates:\n{candidates_text}")
414
+
result = await agent.run(
415
+
f"full tag inventory ({len(tags)} tags after merges):\n{inventory}"
416
+
)
421
417
return [r.model_dump() for r in result.output.relationships]
422
418
423
419