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

Improve Step 1 project matching and Step 2 update-first enforcement

Browse Source 
- Rewrite _STEP1_SYSTEM_PROMPT: lower matching threshold (no longer requires
  "clear" match), strongly prefer existing projects over creating new ones,
  use structured id=|name=|status= format with aiSummary for richer context
- Add code-level UUID validation: reject hallucinated ids not in the fetched
  projects list, fall back to "new" instead of creating a bad link
- Rewrite _PROCESSING_SYSTEM_PROMPT: enforce explicit scan-before-create
  process (read existing → search → update if found → create only if not)
  with hard rule against calling create_* without checking existing records

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This commit is contained in:
Roberto Musso
2026-03-20 23:45:29 +01:00
parent 58bc6efd4b
commit e7cdce8287
1 changed files with 77 additions and 36 deletions
Download Patch File Download Diff File Expand all files Collapse all files

113
app/core/agent_runner.py
View File

@@ -102,19 +102,29 @@ _DOMAIN_DESCRIPTIONS: dict[str, str] = {
_STEP1_SYSTEM_PROMPT = """\
You are a file classifier for a freelance project management tool.
Given a file's content and a list of existing projects, your job is to:
1. Identify which project this file belongs to (or "standalone" if none match).
2. Identify which data domains are relevant to extract from this file,
limited to the allowed domains listed below.
Your job is to match a file to an existing project and identify which data domains to extract.
Domain definitions (only consider domains in the allowed list):
{domain_definitions}
## Project matching rules (STRICT — follow in order)
1. Search the file content for any mention of a project name, client name, acronym, or topic
that overlaps with the existing projects listed below.
2. The match does NOT need to be exact — partial name, abbreviation, or topic similarity is enough.
3. STRONGLY PREFER matching an existing project. Only return "new" as an absolute last resort
when the file has zero meaningful connection to any listed project.
4. When in doubt, pick the closest match from the list.
## Response format
Respond ONLY with a JSON object — no markdown, no explanation:
{{"project_id": "<uuid> or standalone", "domains": ["tasks", "notes"]}}
{{"project_id": "<exact id from the list below, or new>", "new_project_name": "<concise 2-5 word name, only when project_id is new>", "domains": ["tasks", "notes"]}}
## Domain definitions (only consider domains in the allowed list)
{domain_definitions}
## Existing projects
Existing projects:
{projects_list}
"""
@@ -123,20 +133,26 @@ Existing projects:
_PROCESSING_SYSTEM_PROMPT = """\
You are a data extraction assistant for a freelance project management tool.
Your task is to read the file content provided and create or update records
using the available tools.
Your task: extract structured data from the file content and persist it using the available tools.
IMPORTANT — update-first rules:
The existing records below are the source of truth.
If an existing record semantically matches the content (by title, topic,
or context), update it instead of creating a duplicate.
Only create a new record when no existing match is found.
Set isAiSuggested=1 on all new records.
## Mandatory process — follow this order for EVERY item you extract
1. READ the existing records listed below for the relevant domain.
2. SEARCH for a match by title, topic, or semantic similarity.
3. If a match exists → call the update_* tool with the existing record's id.
4. If no match exists → call the create_* tool and set isAiSuggested=1.
NEVER call create_* without first checking the existing records.
NEVER duplicate a record that already exists under a different wording.
## Existing records (source of truth)
{existing_context}
Project context: {project_context}
Target domains: {data_types}
## Context
Project: {project_context}
Domains to extract: {data_types}
{custom_prompt_section}
"""
@@ -470,21 +486,27 @@ async def _classify_file(
file_content: str,
projects: list[dict],
config_data_types: list[str],
) -> tuple[str, list[str]]:
) -> tuple[str, list[str], str | None]:
"""Call the LLM to classify a file by project and relevant domains.
Returns ``(project_id_or_"standalone", domains)``.
Falls back to ``("standalone", config_data_types)`` on any error.
Returns ``(project_id_or_"new", domains, new_project_name_or_None)``.
- ``project_id`` is an existing project UUID, or ``"new"`` when no match found.
- ``new_project_name`` is only set when ``project_id == "new"``.
Falls back to ``("new", config_data_types, None)`` on any error.
"""
fallback = ("standalone", list(config_data_types))
fallback: tuple[str, list[str], str | None] = ("new", list(config_data_types), None)
if not file_content.strip():
return fallback
projects_list = "\n".join(
f" - {p.get('name', '')} (id: {p['id']}, status: {p.get('status', '')})"
for p in projects
) or " (none — all files are standalone)"
valid_project_ids = {p["id"] for p in projects}
def _fmt_project(p: dict) -> str:
summary = (p.get("aiSummary") or p.get("ai_summary") or "").strip()
summary_part = f" — {summary[:100]}" if summary else ""
return f" - id={p['id']} | name={p.get('name', '')} | status={p.get('status', '')}{summary_part}"
projects_list = "\n".join(_fmt_project(p) for p in projects) or " (none yet)"
domain_definitions = "\n".join(
f" - {d}: {_DOMAIN_DESCRIPTIONS[d]}"
@@ -510,14 +532,21 @@ async def _classify_file(
if raw.startswith("json"):
raw = raw[4:]
parsed = json.loads(raw.strip())
project_id: str = str(parsed.get("project_id") or "standalone")
raw_project_id: str = str(parsed.get("project_id") or "new")
# Reject hallucinated UUIDs — only accept ids that exist in the fetched list.
project_id = raw_project_id if raw_project_id in valid_project_ids else "new"
new_project_name: str | None = (
str(parsed["new_project_name"]).strip() or None
if project_id == "new" and parsed.get("new_project_name")
else None
)
domains: list[str] = [
d for d in parsed.get("domains", [])
if d in config_data_types
]
if not domains:
domains = list(config_data_types)
return project_id, domains
return project_id, domains, new_project_name
except Exception as exc:
logger.warning(
"agent_runner: step1 classification failed for %r: %s", file_path, exc
@@ -605,9 +634,6 @@ async def run_local_agent(
# ── Code: fetch all projects once ────────────────────────────
projects = await _fetch_projects()
# ── Per-file processing ──────────────────────────────────────
processing_tools = _build_processing_tools(config.data_types)
for file_path in file_paths:
try:
# Read file content via code.
@@ -622,30 +648,43 @@ async def run_local_agent(
items_processed += 1
# Step 1 — classify file.
project_id, domains = await _classify_file(
project_id, domains, new_project_name = await _classify_file(
file_path=file_path,
file_content=file_content,
projects=projects,
config_data_types=config.data_types,
)
logger.info(
"agent_runner: run=%s file=%r → project=%s domains=%s",
"agent_runner: run=%s file=%r → project=%s new_name=%r domains=%s",
run_id,
file_path,
project_id,
new_project_name,
domains,
)
# Step 2 — fetch existing entities for this project + domains.
# When project_id is "new", entities are fetched without a project
# filter; the LLM will create the project and link records to it.
effective_project_id = project_id if project_id != "new" else "standalone"
existing_blocks: list[str] = []
for domain in domains:
rows = await _fetch_domain_entities(domain, project_id)
rows = await _fetch_domain_entities(domain, effective_project_id)
existing_blocks.append(_format_entities_for_context(domain, rows))
existing_context = "\n\n".join(existing_blocks)
if project_id == "standalone":
project_context = "This file is not associated with any existing project."
if project_id == "new":
name_hint = f' Use "{new_project_name}" as the project name.' if new_project_name else ""
project_context = (
f"No existing project matches this file. "
f"Create a new project first using the create_project tool, "
f"then link all extracted records to its id.{name_hint}"
)
# Ensure the LLM has the project tools available.
if "projects" not in domains:
domains = ["projects"] + domains
else:
project_context = (
f"This file belongs to project ID: {project_id}. "
@@ -659,6 +698,8 @@ async def run_local_agent(
custom_prompt_section=custom_section,
)
processing_tools = _build_processing_tools(domains)
result_text = await _run_agent_with_tools(
system_prompt=system_prompt,
user_message=(