my prefect server setup
prefect-metrics.waow.tech
python
orchestration
reframe relationship discovery as thematic clustering
instead of asking the LLM to evaluate pairs (which sonnet found zero of),
ask it to identify thematic clusters of tags, then derive pairwise edges
from cluster membership. this matches how the tags actually relate — they
form thematic groups, not isolated pairs.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
+55
-21
+44
-21
flows/weave.py
+44
-21
flows/weave.py
···
24
24
from prefect.context import TaskRunContext
25
25
from prefect.variables import Variable
26
26
27
-
from mps.phi import TagMerge, TagRelationship
27
+
from mps.phi import TagCluster, TagMerge, TagRelationship
28
28
29
29
PHI_DID = "did:plc:65sucjiel52gefhcdcypynsr"
30
30
TAG_REL_NAMESPACE = "phi-tag-relationships"
···
336
336
# ---------------------------------------------------------------------------
337
337
338
338
339
-
class RelationshipProposal(BaseModel):
340
-
relationships: list[TagRelationship] = Field(default_factory=list)
339
+
class ClusterProposal(BaseModel):
340
+
clusters: list[TagCluster] = Field(default_factory=list)
341
341
342
342
343
343
@task(
···
356
356
api_key: str,
357
357
model_name: str = "claude-sonnet-4-6",
358
358
) -> list[dict[str, Any]]:
359
-
"""Give the LLM the full tag list with co-occurrence context to find relationships."""
359
+
"""Ask the LLM to identify thematic clusters, then derive pairwise edges."""
360
360
tags = [t for t in sorted(tag_info.keys()) if t not in merged_aliases]
361
361
362
-
# precompute co-occurrence hints to give the LLM as context
362
+
# precompute co-occurrence hints
363
363
cooccur_hints: dict[str, list[str]] = defaultdict(list)
364
364
for pair_key, count in cooccurrences.items():
365
365
t1, t2 = pair_key.split("|", 1)
···
393
393
agent = Agent(
394
394
model,
395
395
system_prompt=(
396
-
"you are mapping relationships between tags in phi's memory graph.\n"
396
+
"you are organizing phi's memory tags into thematic clusters.\n"
397
397
"phi is a bluesky bot that remembers conversations and builds knowledge.\n\n"
398
398
"you will receive the full tag inventory with usage counts, sample observations, "
399
-
"and co-occurrence data. your job: identify genuine conceptual relationships.\n\n"
400
-
"relationship types:\n"
401
-
"- RELATED: broadly connected concepts\n"
402
-
"- SUBTOPIC: one is a narrower form of the other\n"
403
-
"- OVERLAPPING: partially shared meaning but distinct\n\n"
399
+
"and co-occurrence data. your job: identify thematic clusters — groups of tags "
400
+
"that belong to the same area of phi's knowledge or experience.\n\n"
401
+
"examples of good clusters:\n"
402
+
"- 'phi's epistemic concerns': epistemology, memory, confabulation, self-attestation\n"
403
+
"- 'technical interests': programming, ai-systems, infrastructure\n"
404
+
"- 'social dynamics': community, trust, collaboration\n\n"
404
405
"rules:\n"
405
-
"- assign confidence 0.0-1.0 based on how strong the connection is\n"
406
-
"- co-occurrence is a signal but not proof — two tags appearing together "
407
-
"might be coincidental\n"
408
-
"- look for thematic clusters, not just pairs\n"
409
-
"- provide brief evidence for each relationship\n"
410
-
"- be selective — only include relationships you're genuinely confident about\n"
411
-
"- skip trivially obvious connections (like a tag co-occurring with itself)"
406
+
"- a tag can appear in multiple clusters (topics overlap)\n"
407
+
"- clusters should have 2-8 tags — small enough to be coherent\n"
408
+
"- assign cohesion 0.0-1.0: how tightly the tags relate to the cluster theme\n"
409
+
"- name each cluster concisely (2-4 words)\n"
410
+
"- describe what ties the tags together\n"
411
+
"- not every tag needs a cluster — singletons with no thematic neighbors are fine to skip\n"
412
+
"- use the sample observations and co-occurrence data to inform your groupings"
412
413
),
413
-
output_type=RelationshipProposal,
414
-
name="tag-relator",
414
+
output_type=ClusterProposal,
415
+
name="tag-clusterer",
415
416
)
416
417
417
418
result = await agent.run(
418
419
f"full tag inventory ({len(tags)} tags after merges):\n{inventory}"
419
420
)
420
-
return [r.model_dump() for r in result.output.relationships]
421
+
422
+
# derive pairwise relationships from cluster membership
423
+
relationships: list[dict[str, Any]] = []
424
+
seen_pairs: set[tuple[str, str]] = set()
425
+
for cluster in result.output.clusters:
426
+
members = [t for t in cluster.tags if t in set(tags)]
427
+
for i, t1 in enumerate(members):
428
+
for t2 in members[i + 1 :]:
429
+
pair = tuple(sorted([t1, t2]))
430
+
if pair in seen_pairs:
431
+
continue
432
+
seen_pairs.add(pair)
433
+
relationships.append(
434
+
TagRelationship(
435
+
tag_a=pair[0],
436
+
tag_b=pair[1],
437
+
relationship_type="related",
438
+
confidence=cluster.cohesion,
439
+
evidence=f"[{cluster.name}] {cluster.description}",
440
+
).model_dump()
441
+
)
442
+
443
+
return relationships
421
444
422
445
423
446
# ---------------------------------------------------------------------------
+11
packages/mps/src/mps/phi.py
+11
packages/mps/src/mps/phi.py
···
26
26
)
27
27
28
28
29
+
class TagCluster(BaseModel):
30
+
"""A thematic grouping of related tags."""
31
+
32
+
name: str = Field(description="short name for the cluster theme")
33
+
description: str = Field(description="what ties these tags together")
34
+
tags: list[str] = Field(description="member tags")
35
+
cohesion: float = Field(
36
+
ge=0.0, le=1.0, description="how tightly related are these tags"
37
+
)
38
+
39
+
29
40
class TagRelationship(BaseModel):
30
41
"""A discovered relationship between two distinct tags."""
31
42