mirror of
https://github.com/lrnselfreliance/wrolpi.git
synced 2026-10-26 19:01:39 -04:00
ai: ground answers in a live library map, budget the context
Response-quality work, all aimed at giving the model enough context to answer accurately: - Every chat's system prompt now carries a live library map (today's date, counts, and the exact names of channels, domains, zims, and inventories, capped and cached 60s). Exact spellings are in context before the first token, and the list_collections round-trip — minutes on slow hardware — is usually unnecessary. - The request is budgeted against the model's real context: oldest turns are trimmed first, the current question is never dropped, and an over-budget turn gets a clear "start a new conversation" error instead of a silently truncated prompt. cache_prompt reuses the KV cache for the stable prefix, which is most of the latency on a Pi. - context_size plumbing fixed: selecting a model adopts its catalog default (the 4B was running at 8k instead of 16k), and llama-server runs --parallel 1 so the single user gets the whole context. - temperature is configurable in ai.yaml (default 0.3 — small models call tools far more reliably near 0.2-0.3 than at server defaults). - One worked tool-use example per mode is prepended to every conversation; small models imitate far better than they obey. - scripts/ai_eval.py scores the assistant against a live WROLPi on mechanical checks (finished, called the right tool first, included a WROLPi link, no external URLs) so prompt changes stop being validated by anecdote. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This commit is contained in:
parent
82e11cbaba
commit
ced78e181f
11 changed files with 502 additions and 7 deletions
|
|
@ -100,10 +100,15 @@ async def test_chat_tool_call_then_answer(async_client, ai_enabled, video_factor
|
|||
assert dict(events)['tool_result'] == dict(tool='search_videos', success=True)
|
||||
assert dict(events)['done'] == dict(content='Found one video.')
|
||||
|
||||
# The second llama call saw the system prompt, the user message, and the tool result.
|
||||
# The second llama call saw the system prompt (with the live library map), the mode's
|
||||
# worked example, the user message, and the tool result.
|
||||
second_call_messages = calls['messages'][1]
|
||||
assert second_call_messages[0]['role'] == 'system'
|
||||
assert 'librarian' in second_call_messages[0]['content']
|
||||
assert 'Today is' in second_call_messages[0]['content']
|
||||
assert "library contains" in second_call_messages[0]['content']
|
||||
example = [m for m in second_call_messages if m.get('content') and 'Bakehouse' in str(m.get('content'))]
|
||||
assert example, 'the research few-shot example is in context'
|
||||
assert second_call_messages[-1]['role'] == 'tool'
|
||||
assert 'pressure canning' in second_call_messages[-1]['content']
|
||||
|
||||
|
|
@ -152,8 +157,9 @@ async def test_chat_invalid_tool_args(async_client, ai_enabled):
|
|||
assert results[0]['success'] is False # invalid JSON arguments
|
||||
assert results[1]['success'] is False # list_disks is not a help-mode tool
|
||||
assert events[-1] == ('done', dict(content='Sorry.'))
|
||||
# The model saw instructive errors.
|
||||
tool_messages = [m for m in calls['messages'][-1] if m['role'] == 'tool']
|
||||
# The model saw instructive errors (skip the few-shot example's tool messages).
|
||||
tool_messages = [m for m in calls['messages'][-1]
|
||||
if m['role'] == 'tool' and not str(m.get('tool_call_id', '')).startswith('ex_')]
|
||||
assert 'not valid JSON' in tool_messages[0]['content']
|
||||
assert 'Unknown tool' in tool_messages[1]['content']
|
||||
|
||||
|
|
@ -187,3 +193,58 @@ async def test_chat_modes_endpoint(async_client):
|
|||
modes = {m['name']: m for m in response.json['modes']}
|
||||
assert set(modes) == {'help', 'research', 'system'}
|
||||
assert modes['research']['suggestions']
|
||||
|
||||
|
||||
def test_trim_history():
|
||||
"""Old turns are dropped first; fixed messages and the current turn are never trimmed."""
|
||||
from modules.ai import chat
|
||||
fixed = [dict(role='system', content='s' * 400), dict(role='user', content='current')]
|
||||
history = [dict(role='user', content=f'old {i} ' + 'x' * 400) for i in range(10)]
|
||||
|
||||
# A huge budget keeps everything.
|
||||
assert chat.trim_history(history, fixed, [], budget=10_000) == history
|
||||
|
||||
# A small budget drops the OLDEST turns.
|
||||
trimmed = chat.trim_history(history, fixed, [], budget=400)
|
||||
assert len(trimmed) < len(history)
|
||||
assert trimmed == history[len(history) - len(trimmed):]
|
||||
|
||||
# A budget the fixed messages alone exceed empties the history without crashing.
|
||||
assert chat.trim_history(history, fixed, [], budget=1) == []
|
||||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_chat_context_full(async_client, ai_enabled):
|
||||
"""A conversation that cannot fit the model context gets a clear error, not a bad answer."""
|
||||
from modules.ai import chat
|
||||
get_ai_config().context_size = 1 # far below the reply reserve
|
||||
|
||||
patch, calls = _llama_script([dict(type='token', content='hi'), dict(type='stop', tool_calls=[])])
|
||||
body = dict(mode='help', messages=[dict(role='user', content='hello')])
|
||||
with patch:
|
||||
request, response = await async_client.post('/api/chat/', content=json.dumps(body))
|
||||
events = _parse_sse(response.text)
|
||||
assert events[-1][0] == 'error'
|
||||
assert 'no longer fits' in events[-1][1]['message']
|
||||
assert calls['count'] == 0 # the model was never called
|
||||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_chat_history_trimmed(async_client, ai_enabled):
|
||||
"""Old turns are dropped to fit the context; the current question always survives."""
|
||||
get_ai_config().context_size = 4_096
|
||||
|
||||
patch, calls = _llama_script([dict(type='token', content='ok'), dict(type='stop', tool_calls=[])])
|
||||
old_turns = []
|
||||
for i in range(6):
|
||||
old_turns += [dict(role='user', content=f'old question {i} ' + 'pad ' * 500),
|
||||
dict(role='assistant', content=f'old answer {i} ' + 'pad ' * 500)]
|
||||
body = dict(mode='help', messages=old_turns + [dict(role='user', content='the current question')])
|
||||
with patch:
|
||||
request, response = await async_client.post('/api/chat/', content=json.dumps(body))
|
||||
assert _parse_sse(response.text)[-1][0] == 'done'
|
||||
|
||||
sent = calls['messages'][0]
|
||||
contents = [str(m.get('content')) for m in sent]
|
||||
assert any('the current question' in c for c in contents)
|
||||
assert not any('old question 0' in c for c in contents) # oldest turn was trimmed
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue