personal memory agent
feat: migrate speaker_attribution pre-hook to template_vars
Return Layer 4 unmatched speaker context through template_vars instead of rewriting the transcript in the pre hook.
Update the speaker attribution prompt to render unmatched_context inline and add coverage for the Layer 4 pre_process path.
+57
-4
+3
-1
talent/speaker_attribution.md
+3
-1
talent/speaker_attribution.md
···
14
14
15
15
$segment_preamble
16
16
17
+
$unmatched_context
18
+
17
19
# Speaker Attribution — Contextual Identification (Layer 4)
18
20
19
21
## Context
···
23
25
- **Layer 2** used structural heuristics (speaker count, meeting metadata) to label non-owner sentences
24
26
- **Layer 3** matched remaining sentences against known voiceprints
25
27
26
-
The sentences listed above under "Unmatched sentences" could not be resolved by these methods. Your task is to identify the speaker for each unmatched sentence using contextual clues from the transcript.
28
+
The unmatched sentences shown above could not be resolved by these methods. Your task is to identify the speaker for each unmatched sentence using contextual clues from the transcript.
27
29
28
30
## Identification Strategies
29
31
+2
-3
talent/speaker_attribution.py
+2
-3
talent/speaker_attribution.py
···
116
116
lines.append(f'- Sentence {sid}: "{text}"')
117
117
lines.append("")
118
118
119
-
original_transcript = context.get("transcript", "")
120
-
modified_transcript = "\n".join(lines) + "\n" + original_transcript
119
+
unmatched_block = "\n".join(lines)
121
120
122
-
return {"meta": meta, "transcript": modified_transcript}
121
+
return {"meta": meta, "template_vars": {"unmatched_context": unmatched_block}}
123
122
124
123
125
124
def post_process(result: str, context: dict) -> str | None:
+52
tests/test_speaker_attribution_hook.py
+52
tests/test_speaker_attribution_hook.py
···
88
88
stub_path = tmp_path / "agents" / "speaker_labels.json"
89
89
assert not stub_path.exists()
90
90
assert result == {"skip_reason": "no_segment_context"}
91
+
92
+
def test_layer4_returns_template_vars(self):
93
+
"""Unmatched sentences return template vars without changing transcript."""
94
+
result = _run_pre_process(
95
+
{
96
+
**CONTEXT,
97
+
"transcript": "some transcript",
98
+
"meta": {},
99
+
},
100
+
None,
101
+
{
102
+
"labels": [
103
+
{
104
+
"sentence_id": 0,
105
+
"speaker": "owner",
106
+
"confidence": "high",
107
+
"method": "embedding",
108
+
},
109
+
{
110
+
"sentence_id": 1,
111
+
"speaker": None,
112
+
"confidence": None,
113
+
"method": None,
114
+
},
115
+
{
116
+
"sentence_id": 2,
117
+
"speaker": None,
118
+
"confidence": None,
119
+
"method": None,
120
+
},
121
+
],
122
+
"unmatched": [1, 2],
123
+
"unmatched_texts": {1: "Hello everyone", 2: "Let me explain"},
124
+
"candidates": ["Alice Johnson", "Bob Smith"],
125
+
"metadata": {"total": 3, "resolved": 1},
126
+
"source": "audio",
127
+
},
128
+
)
129
+
130
+
assert "meta" in result
131
+
assert "attribution_result" in result["meta"]
132
+
assert "template_vars" in result
133
+
assert "unmatched_context" in result["template_vars"]
134
+
assert "transcript" not in result
135
+
136
+
unmatched_context = result["template_vars"]["unmatched_context"]
137
+
assert "Alice Johnson" in unmatched_context
138
+
assert "Bob Smith" in unmatched_context
139
+
assert "Hello everyone" in unmatched_context
140
+
assert "Let me explain" in unmatched_context
141
+
assert "Sentence 1" in unmatched_context
142
+
assert "Sentence 2" in unmatched_context