feat: add suggest_search_terms with three-angle exclude analysis

Replaces NotImplementedError stub with a real LLMRouter-backed implementation
that builds a structured prompt covering blocklist alias expansion, values
misalignment, and role-type filtering, then parses the JSON response into
suggested_titles and suggested_excludes lists.

Moves LLMRouter import to module level so tests can patch it at
scripts.suggest_helpers.LLMRouter.
This commit is contained in:
pyr0ball 2026-03-05 13:15:25 -08:00
parent efe71150e3
commit 5f1c372c0a
2 changed files with 223 additions and 0 deletions

126
scripts/suggest_helpers.py Normal file
View file

@ -0,0 +1,126 @@
"""
LLM-powered suggestion helpers for Settings UI.
Two functions, each makes one LLMRouter call:
- suggest_search_terms: enhanced title + three-angle exclude suggestions
- suggest_resume_keywords: skills/domains/keywords gap analysis
"""
import json
import re
from pathlib import Path
from typing import Any
from scripts.llm_router import LLMRouter
def _load_resume_context(resume_path: Path) -> str:
"""Extract 3 most recent positions from plain_text_resume.yaml as a short summary."""
import yaml
if not resume_path.exists():
return ""
resume = yaml.safe_load(resume_path.read_text()) or {}
lines = []
for exp in (resume.get("experience_details") or [])[:3]:
pos = exp.get("position", "")
co = exp.get("company", "")
skills = ", ".join((exp.get("skills_acquired") or [])[:5])
lines.append(f"- {pos} at {co}: {skills}")
return "\n".join(lines)
def _parse_json(text: str) -> dict[str, Any]:
"""Extract the first JSON object from LLM output. Returns {} on failure."""
m = re.search(r"\{.*\}", text, re.DOTALL)
if m:
try:
return json.loads(m.group())
except Exception:
pass
return {}
def suggest_search_terms(
current_titles: list[str],
resume_path: Path,
blocklist: dict[str, Any],
user_profile: dict[str, Any],
) -> dict:
"""
Suggest additional job titles and exclude keywords.
Three-angle exclude analysis:
A: Blocklist alias expansion (blocked companies/industries keyword variants)
B: Values misalignment (mission preferences industries/culture to avoid)
C: Role-type filter (career summary role types that don't fit)
Returns: {"suggested_titles": [...], "suggested_excludes": [...]}
"""
resume_context = _load_resume_context(resume_path)
titles_str = "\n".join(f"- {t}" for t in current_titles) or "(none yet)"
bl_companies = ", ".join(blocklist.get("companies", [])) or "none"
bl_industries = ", ".join(blocklist.get("industries", [])) or "none"
nda = ", ".join(user_profile.get("nda_companies", [])) or "none"
career_summary = user_profile.get("career_summary", "") or "Not provided"
mission_raw = user_profile.get("mission_preferences", {}) or {}
mission_str = "\n".join(
f" - {k}: {v}" for k, v in mission_raw.items() if v and v.strip()
) or " (none specified)"
prompt = f"""You are helping a job seeker optimise their search configuration.
--- RESUME BACKGROUND ---
{resume_context or "Not provided"}
--- CAREER SUMMARY ---
{career_summary}
--- CURRENT TITLES BEING SEARCHED ---
{titles_str}
--- BLOCKED ENTITIES ---
Companies blocked: {bl_companies}
Industries blocked: {bl_industries}
NDA / confidential employers: {nda}
--- MISSION & VALUES ---
{mission_str}
Provide all four of the following:
1. TITLE SUGGESTIONS
5-8 additional job titles they may be missing: alternative names, adjacent roles, or senior variants of their current titles.
2. EXCLUDE KEYWORDS BLOCKLIST ALIASES
The user has blocked the companies/industries above. Suggest keyword variants that would also catch their aliases, subsidiaries, or related brands.
Example: blocking "Meta" also exclude "facebook", "instagram", "metaverse", "oculus".
3. EXCLUDE KEYWORDS VALUES MISALIGNMENT
Based on the user's mission and values above, suggest industry or culture keywords to exclude.
Examples: "tobacco", "gambling", "fossil fuel", "defense contractor", "MLM", "commission-only", "pyramid".
4. EXCLUDE KEYWORDS ROLE TYPE FILTER
Based on the user's career background, suggest role-type terms that don't match their trajectory.
Examples for a CS/TAM leader: "cold calling", "door to door", "quota-driven", "SDR", "sales development rep".
Return ONLY valid JSON in exactly this format (no extra text):
{{"suggested_titles": ["Title 1", "Title 2"],
"suggested_excludes": ["keyword 1", "keyword 2", "keyword 3"]}}"""
raw = LLMRouter().complete(prompt).strip()
parsed = _parse_json(raw)
return {
"suggested_titles": parsed.get("suggested_titles", []),
"suggested_excludes": parsed.get("suggested_excludes", []),
}
def suggest_resume_keywords(
resume_path: Path,
current_kw: dict[str, list[str]],
) -> dict:
"""
Suggest skills, domains, and keywords not already in the user's resume_keywords.yaml.
Returns: {"skills": [...], "domains": [...], "keywords": [...]}
"""
raise NotImplementedError

View file

@ -0,0 +1,97 @@
"""Tests for scripts/suggest_helpers.py."""
import json
import pytest
from pathlib import Path
from unittest.mock import patch, MagicMock
RESUME_PATH = Path(__file__).parent.parent / "config" / "plain_text_resume.yaml"
# ── _parse_json ───────────────────────────────────────────────────────────────
def test_parse_json_extracts_valid_object():
from scripts.suggest_helpers import _parse_json
raw = 'Here is the result: {"a": [1, 2], "b": "hello"} done.'
assert _parse_json(raw) == {"a": [1, 2], "b": "hello"}
def test_parse_json_returns_empty_on_invalid():
from scripts.suggest_helpers import _parse_json
assert _parse_json("no json here") == {}
assert _parse_json('{"broken": ') == {}
# ── suggest_search_terms ──────────────────────────────────────────────────────
BLOCKLIST = {
"companies": ["Meta", "Amazon"],
"industries": ["gambling"],
"locations": [],
}
USER_PROFILE = {
"career_summary": "Customer success leader with 10 years in B2B SaaS.",
"mission_preferences": {
"animal_welfare": "I volunteer at my local shelter.",
"education": "",
},
"nda_companies": ["Acme Corp"],
}
def _mock_llm(response_dict: dict):
"""Return a patcher that makes LLMRouter().complete() return a JSON string."""
mock_router = MagicMock()
mock_router.complete.return_value = json.dumps(response_dict)
return patch("scripts.suggest_helpers.LLMRouter", return_value=mock_router)
def test_suggest_search_terms_returns_titles_and_excludes():
from scripts.suggest_helpers import suggest_search_terms
payload = {"suggested_titles": ["VP Customer Success"], "suggested_excludes": ["cold calling"]}
with _mock_llm(payload):
result = suggest_search_terms(["Customer Success Manager"], RESUME_PATH, BLOCKLIST, USER_PROFILE)
assert result["suggested_titles"] == ["VP Customer Success"]
assert result["suggested_excludes"] == ["cold calling"]
def test_suggest_search_terms_prompt_contains_blocklist_companies():
from scripts.suggest_helpers import suggest_search_terms
with _mock_llm({"suggested_titles": [], "suggested_excludes": []}) as mock_cls:
suggest_search_terms(["CSM"], RESUME_PATH, BLOCKLIST, USER_PROFILE)
prompt_sent = mock_cls.return_value.complete.call_args[0][0]
assert "Meta" in prompt_sent
assert "Amazon" in prompt_sent
def test_suggest_search_terms_prompt_contains_mission():
from scripts.suggest_helpers import suggest_search_terms
with _mock_llm({"suggested_titles": [], "suggested_excludes": []}) as mock_cls:
suggest_search_terms(["CSM"], RESUME_PATH, BLOCKLIST, USER_PROFILE)
prompt_sent = mock_cls.return_value.complete.call_args[0][0]
assert "animal_welfare" in prompt_sent or "animal welfare" in prompt_sent.lower()
def test_suggest_search_terms_prompt_contains_career_summary():
from scripts.suggest_helpers import suggest_search_terms
with _mock_llm({"suggested_titles": [], "suggested_excludes": []}) as mock_cls:
suggest_search_terms(["CSM"], RESUME_PATH, BLOCKLIST, USER_PROFILE)
prompt_sent = mock_cls.return_value.complete.call_args[0][0]
assert "Customer success leader" in prompt_sent
def test_suggest_search_terms_returns_empty_on_bad_json():
from scripts.suggest_helpers import suggest_search_terms
mock_router = MagicMock()
mock_router.complete.return_value = "sorry, I cannot help with that"
with patch("scripts.suggest_helpers.LLMRouter", return_value=mock_router):
result = suggest_search_terms(["CSM"], RESUME_PATH, BLOCKLIST, USER_PROFILE)
assert result == {"suggested_titles": [], "suggested_excludes": []}
def test_suggest_search_terms_raises_on_llm_exhausted():
from scripts.suggest_helpers import suggest_search_terms
mock_router = MagicMock()
mock_router.complete.side_effect = RuntimeError("All LLM backends exhausted")
with patch("scripts.suggest_helpers.LLMRouter", return_value=mock_router):
with pytest.raises(RuntimeError, match="All LLM backends exhausted"):
suggest_search_terms(["CSM"], RESUME_PATH, BLOCKLIST, USER_PROFILE)