feat(projekt-matching): vLLM guided-json client with extraction and matching prompts
This commit is contained in:
177
projekt-matching/projektmatch/llm.py
Normal file
177
projekt-matching/projektmatch/llm.py
Normal file
@@ -0,0 +1,177 @@
|
|||||||
|
"""vLLM (OpenAI-compatible) calls with guided_json + German prompts.
|
||||||
|
|
||||||
|
The model is a REASONING model: max_tokens stays UNSET so long thinking
|
||||||
|
chains cannot truncate the final answer (65k context window).
|
||||||
|
"""
|
||||||
|
from __future__ import annotations
|
||||||
|
|
||||||
|
import json
|
||||||
|
import re
|
||||||
|
|
||||||
|
import requests
|
||||||
|
|
||||||
|
THINK_RE = re.compile(r"<think>.*?</think>", re.S)
|
||||||
|
PAGE_TEXT_LIMIT = 24000
|
||||||
|
|
||||||
|
|
||||||
|
class LlmError(RuntimeError):
|
||||||
|
pass
|
||||||
|
|
||||||
|
|
||||||
|
EXTRACT_SCHEMA = {
|
||||||
|
"type": "object",
|
||||||
|
"properties": {
|
||||||
|
"projectName": {"type": "string"},
|
||||||
|
"offerType": {"enum": ["Projekt", "Arbeitnehmer-Angebot", "ANÜ"]},
|
||||||
|
"buyerType": {"enum": ["agency", "direct"]},
|
||||||
|
"companyName": {"type": ["string", "null"]},
|
||||||
|
"contactPerson": {"type": ["string", "null"]},
|
||||||
|
"requirements": {
|
||||||
|
"type": "array",
|
||||||
|
"items": {
|
||||||
|
"type": "object",
|
||||||
|
"properties": {
|
||||||
|
"text": {"type": "string"},
|
||||||
|
"kat": {"enum": ["Must", "Nice", "Misc"]},
|
||||||
|
"miscType": {"enum": ["start", "workload", "duration",
|
||||||
|
"location", "security", "other"]},
|
||||||
|
"startDate": {"type": ["string", "null"]},
|
||||||
|
"workloadPercent": {"type": ["integer", "null"]},
|
||||||
|
"remotePercent": {"type": ["integer", "null"]},
|
||||||
|
"onsiteLocation": {"type": ["string", "null"]},
|
||||||
|
},
|
||||||
|
"required": ["text", "kat", "miscType", "startDate",
|
||||||
|
"workloadPercent", "remotePercent",
|
||||||
|
"onsiteLocation"],
|
||||||
|
},
|
||||||
|
},
|
||||||
|
},
|
||||||
|
"required": ["projectName", "offerType", "buyerType", "companyName",
|
||||||
|
"contactPerson", "requirements"],
|
||||||
|
}
|
||||||
|
|
||||||
|
MATCH_SCHEMA = {
|
||||||
|
"type": "object",
|
||||||
|
"properties": {
|
||||||
|
"ratings": {
|
||||||
|
"type": "array",
|
||||||
|
"items": {
|
||||||
|
"type": "object",
|
||||||
|
"properties": {
|
||||||
|
"nr": {"type": "integer"},
|
||||||
|
"rating": {"enum": ["yes", "no", "unknown"]},
|
||||||
|
"reason": {"type": "string"},
|
||||||
|
},
|
||||||
|
"required": ["nr", "rating", "reason"],
|
||||||
|
},
|
||||||
|
},
|
||||||
|
},
|
||||||
|
"required": ["ratings"],
|
||||||
|
}
|
||||||
|
|
||||||
|
EXTRACT_SYSTEM = """Du extrahierst Anforderungen aus deutschen \
|
||||||
|
Freiberufler-Projektausschreibungen. Antworte NUR mit JSON nach Schema.
|
||||||
|
|
||||||
|
Regeln:
|
||||||
|
- projectName: Titel der Ausschreibung OHNE Portal-Zusatz (z. B. ohne "auf \
|
||||||
|
www.freelancermap.de").
|
||||||
|
- requirements: jede Anforderung einzeln, im Originalwortlaut (behutsames \
|
||||||
|
Kürzen erlaubt, Bedeutung nie verändern). Rahmenbedingungen (Start, \
|
||||||
|
Einsatzort/Remote-Anteil, Auslastung, Laufzeit) sind Anforderungen der \
|
||||||
|
Kategorie Misc.
|
||||||
|
- kat: Must = zwingend (Abschnitt "Must-haves"/"Anforderungen"; "zwingend", \
|
||||||
|
"erforderlich", "vorausgesetzt", "sehr gute Kenntnisse"). Nice = optional \
|
||||||
|
(Abschnitt "Nice-to-haves"; "von Vorteil", "wünschenswert", "idealerweise", \
|
||||||
|
"plus"). Misc = Rahmenbedingungen und alles, was weder Muss noch \
|
||||||
|
Wunsch-Qualifikation ist. Explizite Abschnittsüberschriften haben Vorrang \
|
||||||
|
vor Signalwörtern; "idealerweise" INNERHALB einer Must-Zeile lässt die \
|
||||||
|
Zeile Must bleiben.
|
||||||
|
- miscType nur für Misc-Zeilen relevant (sonst "other"): start = \
|
||||||
|
Projektstart/Verfügbarkeit (startDate als ISO-Datum YYYY-MM-DD, wenn ein \
|
||||||
|
konkretes Datum genannt ist, sonst null); workload = Auslastung \
|
||||||
|
(workloadPercent 0-100 oder null); duration = Laufzeit; location = \
|
||||||
|
Einsatzort/Remote (remotePercent 0-100 oder null; onsiteLocation = \
|
||||||
|
Ortsname oder null); security = Sicherheitsüberprüfung (SÜ, SÜ1/SÜ2/SÜ3, \
|
||||||
|
Ü2, Geheimschutz).
|
||||||
|
- offerType: "Projekt" = Freiberufler-/Werkauftrag (auch über Agentur). \
|
||||||
|
"Arbeitnehmer-Angebot" bei Festanstellung ("Festanstellung", "unbefristet", \
|
||||||
|
"Gehalt", "Arbeitsvertrag"). "ANÜ" bei Arbeitnehmerüberlassung ("ANÜ", \
|
||||||
|
"AÜG", "Überlassung", "Zeitarbeit"). Im Zweifel "Projekt".
|
||||||
|
- buyerType: "agency" bei Personaldienstleistern/Vermittlern (Hays, \
|
||||||
|
GULP/Randstad, SThree, Computer Futures, Aristo, freelancermap-Vermittler, \
|
||||||
|
"im Auftrag unseres Kunden", "für unseren Kunden"), sonst "direct". Im \
|
||||||
|
Zweifel "agency".
|
||||||
|
- companyName: Name der Agentur bzw. des Endkunden, sonst null. \
|
||||||
|
contactPerson: vollständiger Name der Ansprechperson, sonst null."""
|
||||||
|
|
||||||
|
MATCH_SYSTEM = """Du bewertest nüchtern und streng, ob ein Lebenslauf \
|
||||||
|
einzelne Projekt-Anforderungen abdeckt. Antworte NUR mit JSON nach Schema: \
|
||||||
|
für JEDE übergebene Nr. genau ein Eintrag in ratings.
|
||||||
|
|
||||||
|
Bewertung:
|
||||||
|
- "yes" NUR bei klarer Evidenz im Lebenslauf.
|
||||||
|
- "no" wenn der Lebenslauf nichts Belastbares hergibt. Streng bleiben — \
|
||||||
|
eine geschönte Bewertung macht die Match-Zahlen wertlos. Kalibrierung: \
|
||||||
|
Proof-of-Concept-Erfahrung deckt "produktiven Betrieb" NICHT ab; \
|
||||||
|
"mehrjährig" wörtlich nehmen; ein Produktname (z. B. "Azure DevOps \
|
||||||
|
Server") belegt KEINE Cloud-Plattform-Erfahrung.
|
||||||
|
- "unknown" NUR bei echter Teilevidenz, wenn die Entscheidung von Wissen \
|
||||||
|
abhängt, das nur der Kandidat selbst hat.
|
||||||
|
- "wie z. B."-Aufzählungen: gleichwertige Alternativen zählen als Abdeckung \
|
||||||
|
(Beispiel: Ollama/llama.cpp/Transformers decken "LLM-Inference-Stacks wie \
|
||||||
|
z. B. vLLM, TGI, Triton" ab).
|
||||||
|
- reason: EIN kurzer deutscher Satz mit der Begründung."""
|
||||||
|
|
||||||
|
|
||||||
|
def _post(base, model, messages, body_extra):
|
||||||
|
body = {"model": model, "messages": messages, "temperature": 0.1}
|
||||||
|
body.update(body_extra)
|
||||||
|
return requests.post(f"{base.rstrip('/')}/chat/completions", json=body,
|
||||||
|
timeout=1500)
|
||||||
|
|
||||||
|
|
||||||
|
def chat_json(base, model, messages, schema):
|
||||||
|
resp = _post(base, model, messages, {"guided_json": schema})
|
||||||
|
if resp.status_code == 400 and "guided_json" in getattr(resp, "text", ""):
|
||||||
|
resp = _post(base, model, messages, {"response_format": {
|
||||||
|
"type": "json_schema",
|
||||||
|
"json_schema": {"name": "out", "schema": schema}}})
|
||||||
|
if resp.status_code != 200:
|
||||||
|
raise LlmError(f"vLLM HTTP {resp.status_code}: "
|
||||||
|
f"{getattr(resp, 'text', '')[:300]}")
|
||||||
|
content = resp.json()["choices"][0]["message"]["content"] or ""
|
||||||
|
content = THINK_RE.sub("", content).strip()
|
||||||
|
try:
|
||||||
|
return json.loads(content)
|
||||||
|
except json.JSONDecodeError as exc:
|
||||||
|
raise LlmError(f"LLM lieferte kein JSON: {exc}: {content[:200]}")
|
||||||
|
|
||||||
|
|
||||||
|
def extract_project(base, model, page_text):
|
||||||
|
messages = [{"role": "system", "content": EXTRACT_SYSTEM},
|
||||||
|
{"role": "user", "content":
|
||||||
|
"Ausschreibungstext:\n\n" + page_text[:PAGE_TEXT_LIMIT]}]
|
||||||
|
out = chat_json(base, model, messages, EXTRACT_SCHEMA)
|
||||||
|
if not out.get("requirements"):
|
||||||
|
raise LlmError("Extraktion ohne Anforderungen")
|
||||||
|
return out
|
||||||
|
|
||||||
|
|
||||||
|
def match_cv(base, model, cv_text, requirements):
|
||||||
|
listing = "\n".join(f"{r['nr']}. {r['text']}" for r in requirements)
|
||||||
|
messages = [{"role": "system", "content": MATCH_SYSTEM},
|
||||||
|
{"role": "user", "content":
|
||||||
|
f"Lebenslauf:\n\n{cv_text}\n\nAnforderungen:\n{listing}"}]
|
||||||
|
wanted = {r["nr"] for r in requirements}
|
||||||
|
for attempt in range(2):
|
||||||
|
out = chat_json(base, model, messages, MATCH_SCHEMA)
|
||||||
|
got = {r["nr"]: r for r in out.get("ratings", []) if r["nr"] in wanted}
|
||||||
|
if set(got) == wanted:
|
||||||
|
return [got[n] for n in sorted(got)]
|
||||||
|
missing = sorted(wanted - set(got))
|
||||||
|
messages = messages + [
|
||||||
|
{"role": "assistant", "content": json.dumps(out)},
|
||||||
|
{"role": "user", "content":
|
||||||
|
f"Es fehlen Bewertungen für Nr. {missing}. Antworte erneut "
|
||||||
|
f"mit ratings für ALLE Nummern."}]
|
||||||
|
raise LlmError(f"Matching unvollständig, fehlend: {missing}")
|
||||||
58
projekt-matching/tests/test_llm.py
Normal file
58
projekt-matching/tests/test_llm.py
Normal file
@@ -0,0 +1,58 @@
|
|||||||
|
import json
|
||||||
|
from unittest import mock
|
||||||
|
|
||||||
|
import pytest
|
||||||
|
|
||||||
|
from projektmatch import llm
|
||||||
|
|
||||||
|
|
||||||
|
def fake_post(payloads):
|
||||||
|
"""Return a mock for requests.post yielding chat completions."""
|
||||||
|
responses = []
|
||||||
|
for p in payloads:
|
||||||
|
r = mock.Mock()
|
||||||
|
r.status_code = 200
|
||||||
|
r.json.return_value = {"choices": [{"message": {"content": p}}]}
|
||||||
|
responses.append(r)
|
||||||
|
return mock.Mock(side_effect=responses)
|
||||||
|
|
||||||
|
|
||||||
|
def test_chat_json_strips_think_and_parses():
|
||||||
|
content = "<think>lange Kette</think>{\"a\": 1}"
|
||||||
|
with mock.patch("projektmatch.llm.requests.post", fake_post([content])) as p:
|
||||||
|
out = llm.chat_json("http://v/v1", "m", [{"role": "user", "content": "x"}],
|
||||||
|
{"type": "object"})
|
||||||
|
assert out == {"a": 1}
|
||||||
|
body = p.call_args.kwargs["json"]
|
||||||
|
assert "max_tokens" not in body
|
||||||
|
assert body["guided_json"] == {"type": "object"}
|
||||||
|
|
||||||
|
|
||||||
|
def test_chat_json_fallback_to_response_format():
|
||||||
|
bad = mock.Mock(status_code=400,
|
||||||
|
text="Unknown parameter: 'guided_json'")
|
||||||
|
good = mock.Mock(status_code=200)
|
||||||
|
good.json.return_value = {"choices": [{"message": {"content": "{}"}}]}
|
||||||
|
with mock.patch("projektmatch.llm.requests.post",
|
||||||
|
mock.Mock(side_effect=[bad, good])) as p:
|
||||||
|
assert llm.chat_json("http://v/v1", "m", [], {"type": "object"}) == {}
|
||||||
|
assert "response_format" in p.call_args.kwargs["json"]
|
||||||
|
|
||||||
|
|
||||||
|
def test_match_cv_retries_on_missing_nr_then_raises():
|
||||||
|
reqs = [{"nr": 1, "text": "Python"}, {"nr": 2, "text": "K8s"}]
|
||||||
|
partial = json.dumps({"ratings": [
|
||||||
|
{"nr": 1, "rating": "yes", "reason": "ok"}]})
|
||||||
|
with mock.patch("projektmatch.llm.requests.post",
|
||||||
|
fake_post([partial, partial])):
|
||||||
|
with pytest.raises(llm.LlmError, match="unvollständig"):
|
||||||
|
llm.match_cv("http://v/v1", "m", "CV", reqs)
|
||||||
|
|
||||||
|
|
||||||
|
def test_match_cv_ok():
|
||||||
|
reqs = [{"nr": 1, "text": "Python"}]
|
||||||
|
full = json.dumps({"ratings": [
|
||||||
|
{"nr": 1, "rating": "unknown", "reason": "Teilevidenz"}]})
|
||||||
|
with mock.patch("projektmatch.llm.requests.post", fake_post([full])):
|
||||||
|
out = llm.match_cv("http://v/v1", "m", "CV", reqs)
|
||||||
|
assert out[0]["rating"] == "unknown"
|
||||||
Reference in New Issue
Block a user