Skip to content

Speakr is a personal, self-hosted web application designed for transcribing audio recordings

Python

fix: carry citation-numbering and research notes as system messages instead of inside assistant history content, so the model can't imitate them into its answers; renderer strips any leaked note as backstop

Murtaza Nasir205343b8afd4
diff --git a/src/services/inquire_agent.py b/src/services/inquire_agent.py
index b31442ff..b52a0c64 100644
--- a/src/services/inquire_agent.py
+++ b/src/services/inquire_agent.py
@@ -606,11 +606,15 @@ KEEP_VERBATIM_MESSAGES = 8 # last 4 turns
def sanitize_history(message_history):
- """Client history -> [{role, content}] with activity provenance folded in.
-
- Assistant messages also carry the citation numbering the UI actually
- displayed (the client assigns numbers at render time), so a follow-up
- like "tell me more about 5" resolves to what the user saw on screen.
+ """Client history -> [{role, content, note?}].
+
+ Assistant messages carry provenance out-of-band in ``note`` (research
+ performed, plus the citation numbering the UI actually displayed, so a
+ follow-up like "tell me more about 5" resolves to what the user saw).
+ The note is injected as a SEPARATE system message at build time, never
+ folded into the assistant's own content: models imitate the tail of
+ their previous answers, and an inline annotation ends up reproduced
+ verbatim at the bottom of new answers.
"""
out = []
for msg in message_history or []:
@@ -620,9 +624,11 @@ def sanitize_history(message_history):
content = msg.get('content')
if role not in ('user', 'assistant') or not isinstance(content, str) or not content.strip():
continue
+ entry = {'role': role, 'content': content}
if role == 'assistant':
+ note_bits = []
if msg.get('activity'):
- content = f"{content}\n\n[research performed for this answer: {str(msg['activity'])[:300]}]"
+ note_bits.append(f"research performed: {str(msg['activity'])[:300]}")
sources = msg.get('sources')
if isinstance(sources, list) and sources:
lines = []
@@ -637,15 +643,18 @@ def sanitize_history(message_history):
title = str(s.get('title') or '')[:120]
lines.append(f"{n} = {title} (recording {rid})")
if lines:
- content = (f"{content}\n\n[citations in this answer were displayed to the "
- f"user numbered: {'; '.join(lines)}]")
- out.append({'role': role, 'content': content})
+ note_bits.append(f"citations were displayed to the user numbered: {'; '.join(lines)}")
+ if note_bits:
+ entry['note'] = ("Note about the assistant's previous answer (system record, "
+ "never reproduce in answers): " + " | ".join(note_bits))
+ out.append(entry)
return out
def estimate_context_tokens(system_prompt, memory, history, toolbox_schemas):
total = _estimate_tokens(system_prompt) + _estimate_tokens(memory or '')
- total += sum(_estimate_tokens(m['content']) for m in history)
+ total += sum(_estimate_tokens(m['content']) + _estimate_tokens(m.get('note', ''))
+ for m in history)
total += _estimate_tokens(json.dumps(toolbox_schemas))
return total
@@ -660,7 +669,10 @@ def compact_history(user_id, memory, history, keep_last=KEEP_VERBATIM_MESSAGES):
return memory, history, 0
to_fold = history[:-keep_last]
kept = history[-keep_last:]
- transcript = "\n\n".join(f"{m['role'].upper()}: {m['content']}" for m in to_fold)
+ transcript = "\n\n".join(
+ f"{m['role'].upper()}: {m['content']}"
+ + (f"\n[{m['note']}]" if m.get('note') else '')
+ for m in to_fold)
prompt = f"""Summarize this conversation-so-far into a compact memory block that lets an assistant \
continue the conversation seamlessly. Capture: topics discussed, questions asked and their answers (with \
any recording titles/ids referenced), decisions or conclusions, and open threads. Be dense and factual; \
@@ -822,7 +834,12 @@ def run_inquire_agent(app, user_id, user_ctx, user_message, message_history,
if memory:
msgs.append({'role': 'system',
'content': f"Conversation memory (earlier turns, summarized):\n{memory}"})
- msgs.extend(history)
+ for m in history:
+ msgs.append({'role': m['role'], 'content': m['content']})
+ # Provenance notes travel as system messages so the model
+ # never mistakes them for part of its own answer style.
+ if m.get('note'):
+ msgs.append({'role': 'system', 'content': m['note']})
msgs.append({'role': 'user', 'content': user_message})
return msgs
diff --git a/templates/inquire.html b/templates/inquire.html
index a7b02a25..38b3c21d 100644
--- a/templates/inquire.html
+++ b/templates/inquire.html
@@ -985,6 +985,15 @@
// this mapping is sent back with the history.
message.sources = [...refs.entries()].map(([id, e]) => ({ n: e.n, recording_id: Number(id), title: e.title }));
}
+ // Backstop: strip leaked system-record notes if the
+ // model ever echoes them ("[citations in this answer
+ // were displayed...]", "[research performed...]").
+ [...tpl.content.querySelectorAll('p')].forEach(p => {
+ const txt = p.textContent.trim();
+ if (/^\[(citations in this answer|research performed|note about the assistant)/i.test(txt) && txt.endsWith(']')) {
+ p.remove();
+ }
+ });
// Strip a model-written trailing Sources/References
// section (the interface appends its own). Removed
// ONLY when everything after the heading is bare
diff --git a/tests/test_inquire_agent.py b/tests/test_inquire_agent.py
index d5903ec1..5c0e9af4 100644
--- a/tests/test_inquire_agent.py
+++ b/tests/test_inquire_agent.py
@@ -489,14 +489,16 @@ def test_sanitize_history_folds_activity_and_drops_junk():
]
out = ia.sanitize_history(history)
assert len(out) == 2
- assert out[1]['content'].startswith('a1')
- assert 'research performed' in out[1]['content']
- assert all(set(m) == {'role', 'content'} for m in out)
+ # Provenance goes into the out-of-band note, NEVER into the assistant's
+ # own content (models imitate their previous answers' tails).
+ assert out[1]['content'] == 'a1'
+ assert 'research performed' in out[1]['note']
-def test_sanitize_history_folds_displayed_citation_numbering():
+def test_sanitize_history_carries_displayed_citation_numbering_as_note():
"""The UI numbers citations at render time; the mapping must reach the
- model so 'tell me more about 5' resolves to what the user saw."""
+ model (as a system-side note) so 'tell me more about 5' resolves to what
+ the user saw — without polluting the assistant content it would mimic."""
history = [
{'role': 'assistant', 'content': 'answer text',
'sources': [{'n': 1, 'recording_id': 58, 'title': 'Integrating AI'},
@@ -505,12 +507,32 @@ def test_sanitize_history_folds_displayed_citation_numbering():
{'role': 'user', 'content': 'tell me more about 2'},
]
out = ia.sanitize_history(history)
- folded = out[0]['content']
- assert 'displayed to the user numbered' in folded
- assert '1 = Integrating AI (recording 58)' in folded
- assert '2 = Impact of AI (recording 73)' in folded
- assert 'dropped' not in folded
- assert set(out[0]) == {'role', 'content'}
+ assert out[0]['content'] == 'answer text'
+ note = out[0]['note']
+ assert 'displayed to the user numbered' in note
+ assert '1 = Integrating AI (recording 58)' in note
+ assert '2 = Impact of AI (recording 73)' in note
+ assert 'dropped' not in note
+
+
+def test_history_notes_become_system_messages_in_the_loop(data):
+ os.environ['INQUIRE_AGENT_TOOL_MODE'] = 'prompt'
+ history = [
+ {'role': 'user', 'content': 'earlier question'},
+ {'role': 'assistant', 'content': 'earlier answer', 'activity': '2 searches',
+ 'sources': [{'n': 1, 'recording_id': data['r1'], 'title': 'Weekly Sync'}]},
+ ]
+ fake = FakeLLM([nonstream_response('follow-up answer')])
+ events = _run(data, fake, history=history)
+ msgs = fake.calls[0]['messages']
+ assistant_msgs = [m for m in msgs if m['role'] == 'assistant']
+ assert assistant_msgs[0]['content'] == 'earlier answer'
+ notes = [m for m in msgs if m['role'] == 'system' and 'displayed to the user numbered' in m['content']]
+ assert len(notes) == 1
+ # The note follows its assistant message
+ idx = msgs.index(assistant_msgs[0])
+ assert msgs[idx + 1] == notes[0]
+ assert any(e.get('delta') == 'follow-up answer' for e in events)
def test_compact_history_uses_llm_and_keeps_tail(data, monkeypatch):