adiuvAI/api
1
0
Fork 0
Code Issues 1 Pull Requests Actions Packages Projects Releases Wiki Activity

refactor(contextual): drop floating WS frame, runner, and prompt fallback

Browse Source 
contextual_request + contextual_scope_update are the only WS
flows for ad-hoc contextual chat now. Floating system prompt
constant removed; Langfuse 'floating_system' is deleted in a
separate manual step. Also removes floating-agent LLM slot from
llm.py and the associated LLM_MODEL_FLOATING_AGENT setting entry.
This commit is contained in:
Roberto
2026-05-15 18:53:01 +02:00
parent d63fd5f3b9
commit 052c7e3741
4 changed files with 6 additions and 447 deletions
Download Patch File Download Diff File Expand all files Collapse all files

372
app/core/deep_agent.py
View File

@@ -1,4 +1,4 @@
"""Single-agent runners for home and floating chat contexts."""
"""Single-agent runners for home and contextual chat contexts."""
from __future__ import annotations
@@ -7,7 +7,7 @@ import logging
import re
from datetime import date
from collections.abc import AsyncGenerator
from typing import Any, Literal
from typing import Any
from langchain_core.messages import AIMessage, HumanMessage, SystemMessage, ToolMessage
from langchain_core.tools import tool
@@ -29,9 +29,6 @@ logger = logging.getLogger(__name__)
MAX_HISTORY_TURNS = 20
FloatingDomainType = Literal["task", "timeline", "project", "node"]
FloatingDomainSection = Literal["task", "timeline", "note"]
# Mapping of core-memory language values to natural-language names for prompts.
_LANGUAGE_NAMES: dict[str, str] = {
"en": "English", "it": "Italian", "es": "Spanish",
@@ -354,44 +351,6 @@ For "today" / "tomorrow" queries, prefer list_tasks_due_today / list_timelines_t
{request_context}\
"""
_FLOATING_SYSTEM_PROMPT = """\
You are adiuvAI's floating executive assistant.{user_identity}
You are pinned to a specific entity (task, timeline event, project, or note) and you stay strictly within that scope.
Be a proactive partner: anticipate the next useful action and close with a concrete suggestion or a clarifying question — but stay terse, one short paragraph at most.
# How you work
- Use tools before answering anything factual. Never guess.
- Stay in the floating scope (see Request context). If the user asks something outside scope, answer briefly and suggest opening the home assistant.
- Match the user's tone preference. Default to warm-but-direct.
- When the user asks to remember, forget, or update something, use memory tools.
# Filter discipline
- Never set the `assignee` filter on list_tasks/count_tasks unless the user explicitly names a person ("Marco's tasks") or refers to themselves ("my tasks", "assigned to me", "mine").
- The user's own name in the User profile block is for context only — it is NOT a default filter.
- When in doubt, omit `assignee` and return the global result.
# Output format
Plain text only. Do NOT output XML/HTML-like tags such as <task>, <project>, <note>, <timeline>, or any bracketed-id wrappers, and do NOT output <chart> blocks — those are for the home assistant.
# Date filtering
{date_context}
When filtering by date, take dueDateFrom / dueDateTo (ms epoch UTC) verbatim from the DATE CONTEXT boundary table above. Do NOT compute boundaries from now_ms yourself.
For specific dates not listed, compute local-midnight in the user timezone and convert to UTC ms.
# Language
{language_instruction}
# Known people & projects
{relational_memory}
# Behavioral hints
{proactive_hints}
# Request context
{request_context}\
"""
_CONTEXTUAL_SYSTEM_PROMPT = """You are adiuvAI's contextual assistant. The user is working inside the app and has opened a side chat anchored to a specific view ("current view"). Help them act on that view: recap, plan, create entities, answer questions.
Rules:
@@ -486,19 +445,6 @@ Stay terse — your principal is a busy executive.
{request_context}\
"""
_FLOATING_DOMAIN_CLASSIFIER_PROMPT = (
"You are a strict domain classifier for websocket floating requests. "
"Return ONLY a JSON object with keys: type, id, section. "
"Allowed type values: task, timeline, project, node. "
"Allowed section values: task, timeline, note, or null. "
"Rules: infer from user message intent first; do not blindly trust scope.type. "
"If user asks tasks/timeline/notes for a project, set type=project and section accordingly. "
"If project id is unknown but context.resolved_project_id exists, use it as id. "
"If id is unknown, use null. "
"No markdown, no prose, JSON only."
)
def _as_text(content: Any) -> str:
if content is None:
return ""
@@ -727,70 +673,6 @@ def _normalize_tagged_list_lines(text: str, message: str) -> str:
return "\n".join(output_lines)
_GENERIC_TAG_RE = re.compile(r"</?(task|project|note|timeline|chart)>", re.IGNORECASE)
_BRACKETED_ID_RE = re.compile(r"\[(?:[0-9a-fA-F-]{8,}|[A-Za-z0-9_-]{8,})\]")
_FLOATING_EMPTY_FALLBACK = "No results found."
def _strip_floating_markup_fragment(text: str) -> str:
if not text:
return text
cleaned = _GENERIC_TAG_RE.sub("", text)
return _BRACKETED_ID_RE.sub("", cleaned)
def _strip_floating_markup(text: str) -> str:
"""Ensure floating responses stay plain text with no XML-like tag wrappers."""
if not text:
return text
cleaned = _strip_floating_markup_fragment(text)
# Collapse excessive spaces introduced by tag/id removal while preserving lines.
lines = [re.sub(r"[ \t]{2,}", " ", line).strip() for line in cleaned.splitlines()]
return "\n".join(line for line in lines if line)
def _fallback_from_raw_floating_text(raw_text: str) -> str:
fallback = _strip_floating_markup_fragment(raw_text or "")
fallback = re.sub(r"[ \t]{2,}", " ", fallback).strip()
return fallback or _FLOATING_EMPTY_FALLBACK
class _FloatingStreamSanitizer:
"""Streaming sanitizer that removes floating markup without buffering the full answer."""
def __init__(self) -> None:
self._pending = ""
@staticmethod
def _split_safe_boundary(text: str) -> tuple[str, str]:
boundary = len(text)
last_lt = text.rfind("<")
if last_lt != -1 and ">" not in text[last_lt:]:
boundary = min(boundary, last_lt)
last_lb = text.rfind("[")
if last_lb != -1 and "]" not in text[last_lb:]:
boundary = min(boundary, last_lb)
if boundary == len(text):
return text, ""
return text[:boundary], text[boundary:]
def feed(self, chunk: str) -> str:
combined = f"{self._pending}{chunk}"
safe_text, self._pending = self._split_safe_boundary(combined)
return _strip_floating_markup_fragment(safe_text)
def finalize(self) -> str:
# Drop dangling unfinished wrappers at the very end.
tail = re.sub(r"<[^>\n]*$", "", self._pending)
tail = re.sub(r"\[[^\]\n]*$", "", tail)
self._pending = ""
return _strip_floating_markup_fragment(tail)
def _normalize_memory_label(path_or_label: str) -> str:
value = path_or_label.strip()
if value.startswith("/memories/"):
@@ -971,168 +853,6 @@ def _all_tools_for_user(user_id: str, trace_id: str | None) -> list[Any]:
return [*_all_tools(), *_memory_tools(user_id, trace_id)]
def _detect_domain_section(message: str) -> FloatingDomainSection | None:
lowered = message.lower()
if any(keyword in lowered for keyword in ["timeline", "milestone", "release", "schedule"]):
return "timeline"
if any(keyword in lowered for keyword in ["task", "tasks", "todo", "attivit", "azione"]):
return "task"
if any(keyword in lowered for keyword in ["note", "notes", "memo", "document"]):
return "note"
return None
def _normalize_domain_payload(payload: dict[str, Any], fallback_id: str | None) -> dict[str, str | None]:
type_raw = str(payload.get("type") or "").strip().lower()
domain_type: FloatingDomainType = "task"
if type_raw in {"task", "timeline", "project", "node"}:
domain_type = type_raw
id_value = payload.get("id")
domain_id = id_value if isinstance(id_value, str) and id_value.strip() else None
if domain_type == "project" and not domain_id:
domain_id = fallback_id
section_raw = payload.get("section")
section: FloatingDomainSection | None = None
if isinstance(section_raw, str):
section_candidate = section_raw.strip().lower()
if section_candidate in {"task", "timeline", "note"}:
section = section_candidate
if domain_type != "project":
section = None
return {
"type": domain_type,
"id": domain_id,
"section": section,
}
def _parse_json_object(text: str) -> dict[str, Any] | None:
raw = text.strip()
if not raw:
return None
try:
parsed = json.loads(raw)
return parsed if isinstance(parsed, dict) else None
except json.JSONDecodeError:
pass
match = re.search(r"\{.*\}", raw, re.DOTALL)
if not match:
return None
try:
parsed = json.loads(match.group(0))
except json.JSONDecodeError:
return None
return parsed if isinstance(parsed, dict) else None
def _infer_floating_domain_rule_based(message: str, context: dict[str, Any]) -> dict[str, str | None]:
section = _detect_domain_section(message)
scope = context.get("scope") if isinstance(context, dict) else None
resolved_project_id = context.get("resolved_project_id") if isinstance(context, dict) else None
project_id = resolved_project_id if isinstance(resolved_project_id, str) and resolved_project_id else None
if isinstance(scope, dict):
scope_type = str(scope.get("type") or "").strip().lower()
scope_id = scope.get("id")
scope_id_value = scope_id if isinstance(scope_id, str) and scope_id else None
if scope_type in {"task", "tasks"}:
return {"type": "task", "id": scope_id_value, "section": None}
if scope_type in {"project", "projects"}:
project_scope_id = scope_id_value or project_id
return {
"type": "project",
"id": project_scope_id,
"section": section,
}
if scope_type in {"note", "notes"}:
return {
"type": "node",
"id": scope_id_value,
"section": None,
}
if scope_type in {"timeline", "timelines"}:
return {"type": "timeline", "id": scope_id_value, "section": None}
lowered = message.lower()
if any(keyword in lowered for keyword in ["project", "progetto", "client"]) or project_id:
return {
"type": "project",
"id": project_id,
"section": section,
}
if section == "timeline":
return {"type": "timeline", "id": None, "section": None}
if section == "note":
return {"type": "node", "id": None, "section": None}
return {"type": "task", "id": None, "section": None}
async def _infer_floating_domain(message: str, context: dict[str, Any]) -> dict[str, str | None]:
resolved_project_id = context.get("resolved_project_id") if isinstance(context, dict) else None
project_id = resolved_project_id if isinstance(resolved_project_id, str) and resolved_project_id else None
classifier_context = {
"scope": context.get("scope") if isinstance(context.get("scope"), dict) else None,
"resolved_project_id": project_id,
}
try:
llm = get_agent_llm("classifier")
classifier_messages = [
SystemMessage(content=_FLOATING_DOMAIN_CLASSIFIER_PROMPT),
HumanMessage(
content=(
f"Message:\n{message}\n\n"
f"Context:\n{json.dumps(classifier_context, ensure_ascii=True)}"
)
),
]
lf = get_langfuse()
_, classifier_prompt_obj = get_prompt_or_fallback(
"floating_domain_classifier", _FLOATING_DOMAIN_CLASSIFIER_PROMPT
)
# Extract user/session from context for Langfuse attribution
_debug = context.get("_debug") if isinstance(context, dict) else None
_lf_user = (_debug or {}).get("user_id") if isinstance(_debug, dict) else None
_lf_session = (_debug or {}).get("session_id") if isinstance(_debug, dict) else None
with langfuse_context(user_id=_lf_user, session_id=_lf_session):
if lf:
with lf.start_as_current_observation(
as_type="generation",
name="floating-classifier",
model=model_for_agent("classifier"),
prompt=classifier_prompt_obj,
input=classifier_messages,
) as gen:
response = await llm.ainvoke(classifier_messages)
gen.update(output=_as_text(response.content), usage_details=extract_usage(response))
else:
response = await llm.ainvoke(classifier_messages)
parsed = _parse_json_object(_as_text(response.content))
if parsed is not None:
domain = _normalize_domain_payload(parsed, project_id)
logger.info(
"deep_agent: floating_domain_classified type=%s id=%s section=%s",
domain.get("type"),
domain.get("id"),
domain.get("section"),
)
return domain
logger.warning("deep_agent: floating_domain classifier returned non-json output")
except Exception as exc:
logger.warning("deep_agent: floating_domain classifier failed: %s", exc)
return _infer_floating_domain_rule_based(message, context)
def _history_to_messages(history: list[dict[str, str]] | None) -> list[Any]:
if not history:
return []
@@ -1461,25 +1181,6 @@ async def run_home(user_id: str, message: str, context: dict[str, Any]) -> str:
return _normalize_tagged_list_lines(response, message)
async def run_floating(user_id: str, message: str, context: dict[str, Any]) -> tuple[str, dict[str, str | None]]:
prepared_context = await _prepare_context(message, context)
domain = await _infer_floating_domain(message, prepared_context)
system_prompt, langfuse_prompt = _build_system_prompt("floating_system", _FLOATING_SYSTEM_PROMPT, prepared_context)
response = await _run_single_agent(
user_id=user_id,
system_prompt=system_prompt,
message=message,
context=prepared_context,
langfuse_prompt=langfuse_prompt,
agent_name="floating-agent",
conversation_history=context.get("conversation_history"),
)
sanitized = _strip_floating_markup(response)
if not sanitized and response:
sanitized = _fallback_from_raw_floating_text(response)
return sanitized, domain
async def run_home_stream(
user_id: str,
message: str,
@@ -1526,71 +1227,6 @@ async def run_home_stream(
yield "token", normalized
async def run_floating_stream(
user_id: str,
message: str,
context: dict[str, Any],
) -> AsyncGenerator[tuple[str, Any], None]:
prepared_context = await _prepare_context(message, context)
domain = await _infer_floating_domain(message, prepared_context)
yield "floating_domain", domain
brief_mode: bool = bool(context.get("brief_mode"))
briefing_context_text: str = str(context.get("briefing_context") or "").strip()
if brief_mode and briefing_context_text:
# Stage 2: inject briefing as ground truth context.
# Pre-substitute {briefing_context} in the template (handles both Langfuse {{}} and fallback {})
# before compile_prompt sees the remaining standard variables.
template, langfuse_prompt = get_prompt_or_fallback(
"task_brief_followup_system",
_TASK_BRIEF_FOLLOWUP_SYSTEM_PROMPT,
)
system_prompt = compile_prompt(
template, langfuse_prompt,
date_context=_datetime_context_injection(prepared_context).strip(),
language_instruction=_language_instruction(prepared_context).strip(),
user_identity=_user_identity_injection(prepared_context).strip(),
relational_memory=_relational_memory_injection(prepared_context).strip(),
proactive_hints=_proactive_hints_injection(prepared_context).strip(),
request_context=_request_context_block(prepared_context),
briefing_context=briefing_context_text,
)
else:
system_prompt, langfuse_prompt = _build_system_prompt("floating_system", _FLOATING_SYSTEM_PROMPT, prepared_context)
sanitizer = _FloatingStreamSanitizer()
emitted_sanitized = False
raw_chunks: list[str] = []
async for event in _run_single_agent_stream(
user_id=user_id,
system_prompt=system_prompt,
message=message,
context=prepared_context,
langfuse_prompt=langfuse_prompt,
agent_name="floating-agent",
conversation_history=context.get("conversation_history"),
):
event_type, data = event
if event_type != "token":
yield event
continue
raw_chunk = str(data or "")
raw_chunks.append(raw_chunk)
sanitized_chunk = sanitizer.feed(raw_chunk)
if sanitized_chunk:
emitted_sanitized = True
yield "token", sanitized_chunk
tail = sanitizer.finalize()
if tail:
emitted_sanitized = True
yield "token", tail
if not emitted_sanitized and raw_chunks:
yield "token", _fallback_from_raw_floating_text("".join(raw_chunks))
async def run_contextual_stream(
user_id: str,
message: str,
@@ -1599,8 +1235,8 @@ async def run_contextual_stream(
) -> AsyncGenerator[tuple[str, Any], None]:
"""Run the contextual agent for a single user turn.
Mirrors run_floating_stream's plumbing but injects the rendered scope
block into the system prompt and exposes the contextual tool set.
Injects the rendered scope block into the system prompt and exposes
the contextual tool set.
Note-edit tools (propose_note_edit) are intentionally excluded.
*context contract*: callers MUST include ``context["_debug"]["session_id"]``

1
app/core/llm.py
View File

@@ -103,7 +103,6 @@ def get_llm(
_AGENT_MODEL_SETTINGS: dict[str, Callable[[], str]] = {
"classifier": lambda: settings.LLM_MODEL_CLASSIFIER or settings.LLM_MODEL,
"home-agent": lambda: settings.LLM_MODEL_HOME_AGENT or settings.LLM_MODEL,
"floating-agent": lambda: settings.LLM_MODEL_FLOATING_AGENT or settings.LLM_MODEL,
"unified-processor": lambda: settings.LLM_MODEL_UNIFIED_PROCESSOR or settings.LLM_MODEL,
"cloud-processor": lambda: settings.LLM_MODEL_CLOUD_PROCESSOR or settings.LLM_MODEL,
"brief-agent": lambda: settings.LLM_MODEL_BRIEF_AGENT or settings.LLM_MODEL,