cesnimda/jobtrackingapp
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat: improve admin observability and translation-first summaries

Browse Source
This commit is contained in:
cesnimda
2026-03-22 21:37:30 +01:00
parent 8014c1e890
commit 4c49ffb0d6
8 changed files with 585 additions and 261 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tools/summarizer/app.py
+111 -37
View File
@@ -10,12 +10,21 @@ app = FastAPI(title="Local Summarizer")
MODEL_NAME = "sshleifer/distilbart-cnn-12-6"
MAX_INPUT_CHARS = 20000
MAX_CONTEXT_CHARS = 2200
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
model = AutoModelForSeq2SeqLM.from_pretrained(MODEL_NAME)
model.eval()
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)
def _load_runtime():
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
model = AutoModelForSeq2SeqLM.from_pretrained(MODEL_NAME)
model.eval()
has_cuda = torch.cuda.is_available()
device = torch.device("cuda" if has_cuda else "cpu")
model.to(device)
gpu_name = torch.cuda.get_device_name(0) if has_cuda else None
return tokenizer, model, device, has_cuda, gpu_name
tokenizer, model, device, GPU_AVAILABLE, GPU_NAME = _load_runtime()
cache = TTLCache(maxsize=1024, ttl=60 * 60)
@@ -33,7 +42,13 @@ def _key(text: str, max_length: int, min_length: int, top_skills: int) -> str:
@app.get("/health")
async def health():
return {"ok": True, "model": MODEL_NAME}
return {
"ok": True,
"model": MODEL_NAME,
"device": str(device),
"gpu_available": GPU_AVAILABLE,
"gpu_name": GPU_NAME,
}
_TECH = [
@@ -54,6 +69,17 @@ _TECH_PRIORITY = [
]
_MUST_HAVE_HINTS = [
"must have", "required", "requirements", "you have", "you bring", "essential", "we are looking for",
]
_NICE_TO_HAVE_HINTS = [
"nice to have", "bonus", "preferred", "advantageous", "extra plus",
]
_SCREENING_HINTS = [
"experience with", "hands-on", "demonstrated", "proven", "track record", "delivered",
]
def _rank_tech_skills(skills):
ordered = []
seen = set()
@@ -84,7 +110,7 @@ def _extract_bullets(lines, max_items=8):
continue
if re.match(r"^([-*]|\u2022)\s+", s):
s = re.sub(r"^([-*]|\u2022)\s+", "", s).strip()
if 3 <= len(s) <= 200:
if 3 <= len(s) <= 220:
out.append(s)
if len(out) >= max_items:
break
@@ -96,6 +122,7 @@ def _top_keywords(text: str, limit=6):
stop = {
"with", "from", "that", "this", "will", "have", "your", "their", "about", "role", "team", "work",
"experience", "skills", "requirements", "responsibilities", "company", "using", "ability", "years",
"looking", "candidate", "position", "working", "across", "strong", "building", "support",
}
counts = {}
for word in words:
@@ -106,6 +133,27 @@ def _top_keywords(text: str, limit=6):
return [word for word, _ in ordered[:limit]]
def _first_matching_sentences(text: str, hints, limit=3):
sentences = re.split(r"(?<=[.!?])\s+", text)
found = []
for sentence in sentences:
low = sentence.lower()
if any(hint in low for hint in hints):
cleaned = sentence.strip()
if 20 <= len(cleaned) <= 220:
found.append(cleaned)
if len(found) >= limit:
break
return found
def _trim_line(text: str, max_len: int = 140) -> str:
text = re.sub(r"\s+", " ", text).strip(" -•\t")
if len(text) <= max_len:
return text
return text[: max_len - 1].rstrip() + "…"
def _role_focused_excerpt(text: str) -> dict:
cleaned = _strip_html(text)
lines = [ln.strip() for ln in cleaned.splitlines()]
@@ -162,6 +210,11 @@ def _role_focused_excerpt(text: str) -> dict:
responsibilities = any_bullets[:6]
requirements = any_bullets[6:10]
if not requirements:
requirements = [_trim_line(x) for x in _first_matching_sentences(cleaned, _MUST_HAVE_HINTS, limit=4)]
if not nice:
nice = [_trim_line(x) for x in _first_matching_sentences(cleaned, _NICE_TO_HAVE_HINTS, limit=3)]
focused_parts = []
if responsibilities:
focused_parts.append("Responsibilities:\n- " + "\n- ".join(responsibilities))
@@ -169,7 +222,14 @@ def _role_focused_excerpt(text: str) -> dict:
focused_parts.append("Requirements:\n- " + "\n- ".join(requirements))
if nice:
focused_parts.append("Nice to have:\n- " + "\n- ".join(nice))
focused_parts.append("Context:\n" + cleaned[:1500])
focused_parts.append("Context:\n" + cleaned[:MAX_CONTEXT_CHARS])
screen_focus = []
for item in requirements[:4]:
if any(hint in item.lower() for hint in _SCREENING_HINTS) or len(screen_focus) < 2:
screen_focus.append(_trim_line(item))
if not screen_focus:
screen_focus = [_trim_line(x) for x in _first_matching_sentences(cleaned, _SCREENING_HINTS, limit=3)]
return {
"cleaned": cleaned,
@@ -180,6 +240,7 @@ def _role_focused_excerpt(text: str) -> dict:
"tech": tech_found,
"soft": soft_found,
"keywords": _top_keywords(cleaned),
"screen_focus": screen_focus[:3],
}
@@ -193,7 +254,9 @@ def _model_summarize(text: str, max_length: int, min_length: int) -> str:
attention_mask=attention_mask,
max_length=max_length,
min_length=min_length,
num_beams=4,
num_beams=3,
length_penalty=1.0,
no_repeat_ngram_size=3,
early_stopping=True,
)
return tokenizer.decode(outputs[0], skip_special_tokens=True).strip()
@@ -211,53 +274,64 @@ async def summarize(req: SummarizeRequest):
info = _role_focused_excerpt(req.text)
summary = _model_summarize(info["focused_input"], req.max_length, req.min_length)
ranked_tech = []
for t in _rank_tech_skills(info["tech"]):
if t not in ranked_tech:
ranked_tech.append(t)
uniq_soft = []
for s in info["soft"]:
if s not in uniq_soft:
uniq_soft.append(s)
lines = ["Role summary:", summary]
if info["requirements"]:
lines.append("")
lines.append("What the company wants most:")
for x in info["requirements"][:7]:
lines.append(f"- {x}")
for x in info["requirements"][:5]:
lines.append(f"- {_trim_line(x)}")
if ranked_tech:
lines.append("")
lines.append("Top hard skills:")
for skill in ranked_tech[: req.top_skills]:
lines.append(f"- {skill}")
if info["keywords"]:
lines.append("")
lines.append("Keywords to mirror:")
for keyword in info["keywords"][:5]:
lines.append(f"- {keyword}")
if info["responsibilities"]:
lines.append("")
lines.append("What you would be doing:")
for x in info["responsibilities"][:6]:
lines.append(f"- {x}")
for x in info["responsibilities"][:4]:
lines.append(f"- {_trim_line(x)}")
if info["nice"]:
lines.append("")
lines.append("Nice to have:")
for x in info["nice"][:5]:
lines.append(f"- {x}")
for x in info["nice"][:3]:
lines.append(f"- {_trim_line(x)}")
if info["tech"]:
uniq = []
for t in _rank_tech_skills(info["tech"]):
if t not in uniq:
uniq.append(t)
if uniq_soft:
lines.append("")
lines.append("Top hard skills: " + ", ".join(uniq[: req.top_skills]))
if info["soft"]:
uniq_soft = []
for s in info["soft"]:
if s not in uniq_soft:
uniq_soft.append(s)
lines.append("")
lines.append("Relevant soft skills: " + ", ".join(uniq_soft[:8]))
if info["keywords"]:
lines.append("")
lines.append("Key themes: " + ", ".join(info["keywords"][:6]))
lines.append("Relevant soft skills:")
for soft in uniq_soft[:5]:
lines.append(f"- {soft}")
lines.append("")
lines.append("Interview focus:")
if info["requirements"]:
if info["screen_focus"]:
for x in info["screen_focus"]:
lines.append(f"- Be ready to prove: {_trim_line(x)}")
elif info["requirements"]:
for x in info["requirements"][:3]:
lines.append(f"- Prepare examples that demonstrate: {x}")
elif info["tech"]:
for x in _rank_tech_skills(info["tech"])[:3]:
lines.append(f"- Prepare examples that demonstrate: {_trim_line(x)}")
elif ranked_tech:
for x in ranked_tech[:3]:
lines.append(f"- Be ready to explain your hands-on experience with {x}")
else:
lines.append("- Prepare examples showing relevant impact, collaboration, and delivery.")