Eng-Musa commited on
Commit
fa8c299
·
1 Parent(s): 3229f23

update the models routing

Browse files
__pycache__/config.cpython-312.pyc CHANGED
Binary files a/__pycache__/config.cpython-312.pyc and b/__pycache__/config.cpython-312.pyc differ
 
__pycache__/main.cpython-312.pyc CHANGED
Binary files a/__pycache__/main.cpython-312.pyc and b/__pycache__/main.cpython-312.pyc differ
 
config.py CHANGED
@@ -13,13 +13,12 @@ from dotenv import load_dotenv
13
  env_path = Path(__file__).parent / ".env"
14
  load_dotenv(dotenv_path=env_path)
15
 
16
- # ---------------------------------------------------------------------------
17
- # Google Gemini Settings
18
- # ---------------------------------------------------------------------------
19
  GEMINI_API_KEY = os.getenv("GEMINI_API_KEY", "")
20
-
21
- # Default model: Gemini 3.5 Flash-Lite
22
  GEMINI_MODEL = "gemini-3.5-flash-lite"
23
 
24
  MAX_OUTPUT_TOKENS = 4096
25
  TEMPERATURE = 0.1
 
 
 
 
13
  env_path = Path(__file__).parent / ".env"
14
  load_dotenv(dotenv_path=env_path)
15
 
16
+ # Google Gemini
 
 
17
  GEMINI_API_KEY = os.getenv("GEMINI_API_KEY", "")
 
 
18
  GEMINI_MODEL = "gemini-3.5-flash-lite"
19
 
20
  MAX_OUTPUT_TOKENS = 4096
21
  TEMPERATURE = 0.1
22
+
23
+ # LLM failover chain: first provider is used; on failure logs once and tries next
24
+ LLM_PROVIDERS = ["google", "fallback"]
main.py CHANGED
@@ -1,5 +1,6 @@
1
  from __future__ import annotations
2
 
 
3
  import tempfile
4
  import uuid
5
  import httpx
@@ -9,6 +10,8 @@ from typing import Any, List, Optional
9
 
10
  from contextlib import asynccontextmanager
11
 
 
 
12
  from fastapi import FastAPI, File, Form, UploadFile
13
  from fastapi.exceptions import RequestValidationError
14
  from fastapi.responses import JSONResponse
@@ -21,6 +24,8 @@ from services.cv_converter import CVConverter
21
  from services.job_matcher import JobInput, JobMatcher
22
  from services.workers import CvWorker, QueueManager, CvTask
23
  from services.new.cv_processor_v2 import CvProcessorV2
 
 
24
 
25
  # ---------------------------------------------------------------------------
26
  # Configuration
@@ -417,44 +422,73 @@ def batch_match_cv(payload: BatchMatchCVRequest):
417
 
418
 
419
  # ---------------------------------------------------------------------------
420
- # POST /v2/ping-llm — quick diagnostic endpoint to test Gemini LLM & latency
421
  # ---------------------------------------------------------------------------
422
 
423
 
424
  class PingLLMRequest(BaseModel):
425
- prompt: str = Field(default="Say hello in one short sentence.", description="Prompt to send to Gemini")
426
 
427
 
428
  @app.post("/v2/ping-llm", response_model=APIResponse)
429
  def ping_llm(payload: PingLLMRequest = PingLLMRequest()):
430
- """Diagnostic endpoint to ping Gemini API directly and measure latency."""
 
 
 
431
  t_start = datetime.now(timezone.utc)
432
- try:
433
- res = cv_processor_v2.gemini_extractor.ping(prompt=payload.prompt)
434
- t_end = datetime.now(timezone.utc)
435
- duration_ms = int((t_end - t_start).total_seconds() * 1000)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
436
 
437
- return APIResponse(
438
- message="Gemini ping successful",
439
- statusCode=200,
440
- payload={
441
- "model": res.get("model", ""),
442
- "prompt": payload.prompt,
443
- "reply": res.get("reply", ""),
444
- "duration_ms": duration_ms,
445
- "duration_seconds": round(duration_ms / 1000, 2),
446
- "usageMetadata": res.get("usageMetadata", {}),
447
- },
448
- )
449
- except Exception as exc:
450
- return APIResponse(
451
- message=f"Gemini ping error: {exc}",
452
- statusCode=500,
453
- )
 
 
 
 
 
 
 
 
 
454
 
455
 
456
  # ---------------------------------------------------------------------------
457
- # POST /v2/process-cv — LLM-powered extraction (pymupdf4llm + Gemini API)
458
  # ---------------------------------------------------------------------------
459
 
460
 
@@ -462,9 +496,10 @@ def ping_llm(payload: PingLLMRequest = PingLLMRequest()):
462
  async def process_cv_v2(file: UploadFile = File(...)):
463
  """
464
  V2 CV processing endpoint.
465
- Converts the uploaded file to Markdown then extracts all fields using a
466
- local LLM (Phi-3-mini Q4 by default). Returns markdown, structured fields,
467
- and a sections[] array that preserves the original CV heading titles.
 
468
  """
469
  if not file.filename:
470
  return APIResponse(message="No file uploaded.", statusCode=400)
@@ -508,7 +543,8 @@ async def process_cv_v2(file: UploadFile = File(...)):
508
  statusCode=200,
509
  payload={
510
  **payload,
511
- "timing": {"total_ms": total_ms},
 
512
  "processed_at": t_end.isoformat(),
513
  },
514
  )
 
1
  from __future__ import annotations
2
 
3
+ import logging
4
  import tempfile
5
  import uuid
6
  import httpx
 
10
 
11
  from contextlib import asynccontextmanager
12
 
13
+ logger = logging.getLogger(__name__)
14
+
15
  from fastapi import FastAPI, File, Form, UploadFile
16
  from fastapi.exceptions import RequestValidationError
17
  from fastapi.responses import JSONResponse
 
24
  from services.job_matcher import JobInput, JobMatcher
25
  from services.workers import CvWorker, QueueManager, CvTask
26
  from services.new.cv_processor_v2 import CvProcessorV2
27
+ from services.new.google.gemini_extractor import GeminiCvExtractor
28
+ from config import GEMINI_MODEL, LLM_PROVIDERS
29
 
30
  # ---------------------------------------------------------------------------
31
  # Configuration
 
422
 
423
 
424
  # ---------------------------------------------------------------------------
425
+ # POST /v2/ping-llm — quick diagnostic: ping LLM provider with failover chain
426
  # ---------------------------------------------------------------------------
427
 
428
 
429
  class PingLLMRequest(BaseModel):
430
+ prompt: str = Field(default="Say hello in one short sentence.", description="Prompt to send to the LLM")
431
 
432
 
433
  @app.post("/v2/ping-llm", response_model=APIResponse)
434
  def ping_llm(payload: PingLLMRequest = PingLLMRequest()):
435
+ """
436
+ Diagnostic endpoint. Picks the first provider in LLM_PROVIDERS; if it fails
437
+ logs once and tries the next until all exhausted.
438
+ """
439
  t_start = datetime.now(timezone.utc)
440
+ last_error: Optional[Exception] = None
441
+
442
+ for idx, provider in enumerate(LLM_PROVIDERS):
443
+ p = provider.lower().strip()
444
+ try:
445
+ if p == "google":
446
+ extractor = GeminiCvExtractor()
447
+ res = extractor.ping(prompt=payload.prompt)
448
+ model_name = res.get("model", GEMINI_MODEL)
449
+ reply = res.get("reply", "")
450
+ usage = res.get("usageMetadata", {})
451
+ elif p == "fallback":
452
+ model_name = "regex-fallback"
453
+ reply = "Fallback parser (regex chunker) is operational"
454
+ usage = {}
455
+ else:
456
+ logger.warning("Unknown provider '%s' in LLM_PROVIDERS. Skipping...", p)
457
+ continue
458
+
459
+ t_end = datetime.now(timezone.utc)
460
+ duration_ms = int((t_end - t_start).total_seconds() * 1000)
461
 
462
+ return APIResponse(
463
+ message=f"{p} ping successful",
464
+ statusCode=200,
465
+ payload={
466
+ "provider": p,
467
+ "model": model_name,
468
+ "prompt": payload.prompt,
469
+ "reply": reply,
470
+ "duration_ms": duration_ms,
471
+ "duration_seconds": round(duration_ms / 1000, 2),
472
+ "usageMetadata": usage,
473
+ },
474
+ )
475
+ except Exception as exc:
476
+ next_provider = LLM_PROVIDERS[idx + 1] if idx + 1 < len(LLM_PROVIDERS) else None
477
+ logger.warning(
478
+ "Provider '%s' ping failed: %s. %s",
479
+ p, exc,
480
+ f"Trying next provider '{next_provider}'..." if next_provider else "No more providers.",
481
+ )
482
+ last_error = exc
483
+
484
+ return APIResponse(
485
+ message=f"All providers failed. Last error: {last_error}",
486
+ statusCode=500,
487
+ )
488
 
489
 
490
  # ---------------------------------------------------------------------------
491
+ # POST /v2/process-cv — LLM-powered extraction with provider failover
492
  # ---------------------------------------------------------------------------
493
 
494
 
 
496
  async def process_cv_v2(file: UploadFile = File(...)):
497
  """
498
  V2 CV processing endpoint.
499
+ Converts the uploaded file to Markdown then extracts all fields using the
500
+ configured LLM_PROVIDERS failover chain (google → fallback).
501
+ Returns markdown, structured fields, and a sections[] array that preserves
502
+ the original CV heading titles.
503
  """
504
  if not file.filename:
505
  return APIResponse(message="No file uploaded.", statusCode=400)
 
543
  statusCode=200,
544
  payload={
545
  **payload,
546
+ "provider": payload.get("provider", "unknown"),
547
+ "timing": {"total_ms": total_ms},
548
  "processed_at": t_end.isoformat(),
549
  },
550
  )
services/new/__pycache__/cv_processor_v2.cpython-312.pyc CHANGED
Binary files a/services/new/__pycache__/cv_processor_v2.cpython-312.pyc and b/services/new/__pycache__/cv_processor_v2.cpython-312.pyc differ
 
services/new/cv_processor_v2.py CHANGED
@@ -1,4 +1,4 @@
1
- """Orchestrates V2 CV processing: file → markdown → Gemini extraction → payload."""
2
 
3
  from __future__ import annotations
4
 
@@ -7,6 +7,7 @@ import tempfile
7
  from pathlib import Path
8
  from typing import Any, Dict
9
 
 
10
  from services.cv_chunker import chunk_cv
11
  from services.new.cv_converter_v2 import CvConverterV2
12
  from services.new.google.gemini_extractor import GeminiCvExtractor
@@ -21,7 +22,7 @@ logger = logging.getLogger(__name__)
21
 
22
 
23
  class CvProcessorV2:
24
- """Processes CV using Gemini LLM with 3 retries, falling back to regex chunker if needed."""
25
 
26
  def __init__(self) -> None:
27
  self.converter = CvConverterV2()
@@ -29,8 +30,9 @@ class CvProcessorV2:
29
 
30
  def process(self, file_bytes: bytes, filename: str) -> Dict[str, Any]:
31
  """
32
- Convert file → markdown, then extract fields via Gemini (or older chunk_cv on failure).
33
- Returns payload including 'method_used': 'llm' | 'fallback'.
 
34
  """
35
  suffix = Path(filename).suffix.lower() if filename else ".pdf"
36
  tmp_path = None
@@ -45,73 +47,82 @@ class CvProcessorV2:
45
  if not conversion.success:
46
  raise ValueError(conversion.error or "Conversion failed")
47
 
48
- # Step 2 — Attempt Gemini extraction (handles up to 3 internal retries)
49
- method_used = "llm"
50
- extracted = None
51
- try:
52
- extracted = self.gemini_extractor.extract(conversion.markdown)
53
- except Exception as gemini_err:
54
- logger.warning(
55
- "Gemini extraction failed after 3 retries (%s). Falling back to older regex chunk_cv...",
56
- gemini_err,
57
- )
58
- method_used = "fallback"
59
-
60
  warnings = list(conversion.warnings)
61
 
62
- # --- LLM PATH ---
63
- if method_used == "llm" and extracted:
64
- skills_found = set(s.lower() for s in extracted.get("skills", {}).get("items", []))
65
- if not skills_found:
66
- skills_found = _extract_skills(conversion.markdown, DEFAULT_SKILL_ALIASES)
67
- inferred_category = _infer_category(
68
- text=conversion.markdown,
69
- skills=skills_found,
70
- category_skills=DEFAULT_CATEGORY_SKILLS,
71
- )
72
-
73
- return {
74
- "method_used": "llm",
75
- "markdown": extracted.get("markdown", conversion.markdown),
76
- "cv_title": extracted.get("cv_title", ""),
77
- "seniority": extracted.get("seniority", "entry"),
78
- "years_experience": extracted.get("years_experience", 0),
79
- "category": inferred_category,
80
- "completeness_score": extracted.get("completeness_score", 0),
81
- "summary": extracted.get("summary", ""),
82
- "contact": extracted.get("contact", {}),
83
- "links": extracted.get("links", []),
84
- "skills": extracted.get("skills", {}),
85
- "experience": extracted.get("experience", []),
86
- "education": extracted.get("education", []),
87
- "projects": extracted.get("projects", []),
88
- "awards": extracted.get("awards", []),
89
- "raw_sections": extracted.get("raw_sections", []),
90
- "usage_metadata": extracted.get("usage_metadata", {}),
91
- "file_type": conversion.file_type,
92
- "is_scanned": conversion.is_scanned,
93
- "page_count": conversion.page_count,
94
- "warnings": warnings,
95
- }
96
-
97
- # --- FALLBACK PATH (Returns older structure exactly as v1) ---
98
- warnings.append("Gemini service unavailable. Parsed using regex fallback.")
99
- chunks = chunk_cv(conversion.markdown)
100
-
101
- return {
102
- "method_used": "fallback",
103
- "markdown": conversion.markdown,
104
- "cv_title": chunks.get("cv_title", ""),
105
- "seniority": chunks.get("seniority", "entry"),
106
- "years_experience": chunks.get("years_experience", 0),
107
- "category": chunks.get("category", "Other"),
108
- "completeness_score": chunks.get("completeness_score", 0),
109
- "chunks": chunks.get("chunks", {}),
110
- "file_type": conversion.file_type,
111
- "is_scanned": conversion.is_scanned,
112
- "page_count": conversion.page_count,
113
- "warnings": warnings,
114
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
115
 
116
  finally:
117
  if tmp_path is not None:
 
1
+ """Orchestrates V2 CV processing: file → markdown → LLM extraction → payload."""
2
 
3
  from __future__ import annotations
4
 
 
7
  from pathlib import Path
8
  from typing import Any, Dict
9
 
10
+ from config import LLM_PROVIDERS
11
  from services.cv_chunker import chunk_cv
12
  from services.new.cv_converter_v2 import CvConverterV2
13
  from services.new.google.gemini_extractor import GeminiCvExtractor
 
22
 
23
 
24
  class CvProcessorV2:
25
+ """Processes CV using the configured LLM_PROVIDERS failover chain."""
26
 
27
  def __init__(self) -> None:
28
  self.converter = CvConverterV2()
 
30
 
31
  def process(self, file_bytes: bytes, filename: str) -> Dict[str, Any]:
32
  """
33
+ Convert file → markdown, then extract fields using LLM_PROVIDERS in order.
34
+ If a provider fails, logs once and tries the next. Last entry is always fallback.
35
+ Returns payload including 'method_used': 'llm' | 'fallback' and 'provider'.
36
  """
37
  suffix = Path(filename).suffix.lower() if filename else ".pdf"
38
  tmp_path = None
 
47
  if not conversion.success:
48
  raise ValueError(conversion.error or "Conversion failed")
49
 
 
 
 
 
 
 
 
 
 
 
 
 
50
  warnings = list(conversion.warnings)
51
 
52
+ # Step 2 — provider failover chain
53
+ last_error = None
54
+ for idx, provider in enumerate(LLM_PROVIDERS):
55
+ p = provider.lower().strip()
56
+
57
+ # --- GOOGLE (Gemini) ---
58
+ if p == "google":
59
+ try:
60
+ extracted = self.gemini_extractor.extract(conversion.markdown)
61
+ skills_found = set(s.lower() for s in extracted.get("skills", {}).get("items", []))
62
+ if not skills_found:
63
+ skills_found = _extract_skills(conversion.markdown, DEFAULT_SKILL_ALIASES)
64
+ inferred_category = _infer_category(
65
+ text=conversion.markdown,
66
+ skills=skills_found,
67
+ category_skills=DEFAULT_CATEGORY_SKILLS,
68
+ )
69
+ return {
70
+ "method_used": "llm",
71
+ "provider": "google",
72
+ "markdown": extracted.get("markdown", conversion.markdown),
73
+ "cv_title": extracted.get("cv_title", ""),
74
+ "seniority": extracted.get("seniority", "entry"),
75
+ "years_experience": extracted.get("years_experience", 0),
76
+ "category": inferred_category,
77
+ "completeness_score": extracted.get("completeness_score", 0),
78
+ "summary": extracted.get("summary", ""),
79
+ "contact": extracted.get("contact", {}),
80
+ "links": extracted.get("links", []),
81
+ "skills": extracted.get("skills", {}),
82
+ "experience": extracted.get("experience", []),
83
+ "education": extracted.get("education", []),
84
+ "projects": extracted.get("projects", []),
85
+ "awards": extracted.get("awards", []),
86
+ "raw_sections": extracted.get("raw_sections", []),
87
+ "usage_metadata": extracted.get("usage_metadata", {}),
88
+ "file_type": conversion.file_type,
89
+ "is_scanned": conversion.is_scanned,
90
+ "page_count": conversion.page_count,
91
+ "warnings": warnings,
92
+ }
93
+ except Exception as exc:
94
+ next_provider = LLM_PROVIDERS[idx + 1] if idx + 1 < len(LLM_PROVIDERS) else None
95
+ logger.warning(
96
+ "Provider 'google' extraction failed: %s. %s",
97
+ exc,
98
+ f"Trying next provider '{next_provider}'..." if next_provider else "No more providers.",
99
+ )
100
+ last_error = exc
101
+
102
+ # --- FALLBACK (regex chunker) ---
103
+ elif p == "fallback":
104
+ warnings.append("LLM service unavailable. Parsed using regex fallback.")
105
+ chunks = chunk_cv(conversion.markdown)
106
+ return {
107
+ "method_used": "fallback",
108
+ "provider": "fallback",
109
+ "markdown": conversion.markdown,
110
+ "cv_title": chunks.get("cv_title", ""),
111
+ "seniority": chunks.get("seniority", "entry"),
112
+ "years_experience": chunks.get("years_experience", 0),
113
+ "category": chunks.get("category", "Other"),
114
+ "completeness_score": chunks.get("completeness_score", 0),
115
+ "chunks": chunks.get("chunks", {}),
116
+ "file_type": conversion.file_type,
117
+ "is_scanned": conversion.is_scanned,
118
+ "page_count": conversion.page_count,
119
+ "warnings": warnings,
120
+ }
121
+
122
+ else:
123
+ logger.warning("Unknown provider '%s' in LLM_PROVIDERS. Skipping...", p)
124
+
125
+ raise RuntimeError(f"All extraction providers failed. Last error: {last_error}")
126
 
127
  finally:
128
  if tmp_path is not None:
services/new/google/__pycache__/gemini_extractor.cpython-312.pyc CHANGED
Binary files a/services/new/google/__pycache__/gemini_extractor.cpython-312.pyc and b/services/new/google/__pycache__/gemini_extractor.cpython-312.pyc differ