added multiple llm option

README.md

@@ -12,11 +12,11 @@ license: mit
 
 # Smart Resume Builder
 
-Generate grounded, tailored resumes from a job description and a PDF resume using Gradio, OpenAI, and LaTeX templates. Suitable for local runs or Hugging Face Spaces. 
+Generate grounded, tailored resumes from a job description and a PDF resume using Gradio, OpenAI or Hugging Face models, and LaTeX templates. Suitable for local runs or Hugging Face Spaces. 
 
 ## Features
 - PDF parsing with `pdfplumber` and `pymupdf` fallback
-- Strict, evidence-backed JSON extraction via OpenAI
+- Strict, evidence-backed JSON extraction via OpenAI or Hugging Face Inference API
 - Tailoring step that rewrites bullets without fabrication and reports missing items
 - Two LaTeX templates (modern single-column and classic two-column)
 - Streamlit UI with API key storage (keyring preferred), template selector, and export buttons
@@ -33,7 +33,7 @@ uv run app.py
 ## Using the app
 1. Paste the job description.
 2. Upload a resume PDF.
-3. Enter your OpenAI API key (optionally save it locally; system keychain is used when available).
+3. Pick a provider and enter your API key/token (optionally save it locally; system keychain is used when available).
 4. Choose a model name and LaTeX template.
 5. Click **Generate Tailored Resume**.
 6. Review the LaTeX preview, missing/needs-confirmation list, and keyword alignment.
@@ -63,4 +63,4 @@ uv run pytest
 ## Troubleshooting
 - Missing `latexmk`: install TeX Live/MikTeX.
 - If PDF parsing is poor, ensure the resume PDF is text-based; image-only scans are harder to extract.
-- For OpenAI errors, verify the API key and model name in the UI.
+- For provider errors, verify the API key/token and model name in the UI.

app.py

@@ -19,6 +19,13 @@ logger = logging.getLogger("smart_resume_builder")
 
 APP_TITLE = "Smart Resume Builder"
 LOCAL_KEY_PATH = Path.home() / ".smart_resume_builder_key"
+OPENAI_MODELS = ["gpt-4o-mini", "gpt-4o", "gpt-4.1-mini"]
+HF_MODELS = [
+    "mistralai/Mistral-7B-Instruct-v0.2",
+    "mistralai/Mixtral-8x7B-Instruct-v0.1",
+    "HuggingFaceH4/zephyr-7b-beta",
+]
+HF_PROVIDER_LABEL = "Hugging Face (Inference API)"
 
 
 # Gradio 4.44.1 can emit JSON schema fragments with `additionalProperties: true`,
@@ -36,6 +43,12 @@ def _safe_json_schema_to_python_type(schema, defs=None):
 gr_client_utils._json_schema_to_python_type = _safe_json_schema_to_python_type
 
 
+def _provider_defaults(provider: str) -> Tuple[list[str], str, str]:
+    if provider == HF_PROVIDER_LABEL:
+        return HF_MODELS, HF_MODELS[0], "Hugging Face Token"
+    return OPENAI_MODELS, OPENAI_MODELS[0], "OpenAI API Key"
+
+
 def load_api_key() -> Optional[str]:
     try:
         import keyring  # type: ignore
@@ -97,6 +110,7 @@ def generate_tailored_resume(
     job_description: str,
     pdf_file,
     api_key: str,
+    provider: str,
     model: str,
     template_choice: str,
     save_key: bool,
@@ -107,7 +121,16 @@ def generate_tailored_resume(
         logs.append(msg)
 
     if not api_key:
-        return ("", "API key required.", "", {}, "\n".join(logs), None, None, {})
+        return (
+            "",
+            "API key/token required.",
+            "",
+            {},
+            "\n".join(logs),
+            None,
+            None,
+            {},
+        )
     if not pdf_file:
         return ("", "Please upload a resume PDF.", "", {}, "\n".join(logs), None, None, {})
     if not job_description.strip():
@@ -123,7 +146,7 @@ def generate_tailored_resume(
             pdf_path = Path(tmp.name)
         result = parse_resume_pdf(str(pdf_path))
         log(f"Extracted text using {result.method}")
-        log("Starting OpenAI pipeline...")
+        log(f"Starting LLM pipeline (provider={provider})...")
 
         template_map = list_templates()
         template_source = template_map[template_choice].read_text(encoding="utf-8")
@@ -131,6 +154,7 @@ def generate_tailored_resume(
         resume, tailored = run_pipeline(
             api_key=api_key,
             model=model,
+            provider=provider,
             raw_text=result.raw_text,
             job_description=job_description,
             template_name=template_choice,
@@ -173,7 +197,7 @@ def generate_tailored_resume(
     except Exception as exc:
         log(f"Error: {exc}")
         log(
-            "If this persists, verify your OpenAI API key/model and that outbound network access is allowed."
+            "If this persists, verify your API key/token and model and that outbound network access is allowed."
         )
         return (
             "",
@@ -197,9 +221,19 @@ def build_ui():
         with gr.Row():
             with gr.Column():
                 jd = gr.Textbox(label="Job Description", lines=12, placeholder="Paste JD here")
+                provider = gr.Dropdown(
+                    label="Provider",
+                    choices=["OpenAI", HF_PROVIDER_LABEL],
+                    value="OpenAI",
+                )
                 api = gr.Textbox(label="OpenAI API Key", type="password", value=stored_key)
                 save_key = gr.Checkbox(label="Save key locally (keyring preferred)", value=bool(stored_key))
-                model = gr.Textbox(label="Model name", value="gpt-4o-mini")
+                model = gr.Dropdown(
+                    label="Model name",
+                    choices=OPENAI_MODELS,
+                    value=OPENAI_MODELS[0],
+                    allow_custom_value=True,
+                )
                 template_choice = gr.Dropdown(
                     label="Template", choices=template_names, value=template_names[0]
                 )
@@ -219,7 +253,7 @@ def build_ui():
 
         generate_btn.click(
             fn=generate_tailored_resume,
-            inputs=[jd, pdf, api, model, template_choice, save_key],
+            inputs=[jd, pdf, api, provider, model, template_choice, save_key],
             outputs=[
                 latex_preview,
                 missing_panel,
@@ -232,6 +266,19 @@ def build_ui():
             ],
         )
 
+        def _update_provider_fields(selected: str):
+            choices, value, key_label = _provider_defaults(selected)
+            return (
+                gr.Dropdown.update(choices=choices, value=value),
+                gr.Textbox.update(label=key_label),
+            )
+
+        provider.change(
+            fn=_update_provider_fields,
+            inputs=provider,
+            outputs=[model, api],
+        )
+
         clear_btn.click(fn=clear_api_key, inputs=None, outputs=api)
 
     return demo

llm/client.py

@@ -3,11 +3,17 @@ from __future__ import annotations
 import json
 import logging
 import time
-from typing import Any, Dict, List
+from typing import Any, Dict, Protocol
 
 logger = logging.getLogger(__name__)
 
 
+class LLMClient(Protocol):
+    def chat(self, prompt: str, *, max_retries: int = 3) -> str: ...
+
+    def chat_json(self, prompt: str, *, max_retries: int = 3) -> Dict[str, Any]: ...
+
+
 class OpenAIClient:
     def __init__(self, api_key: str, model: str = "gpt-4o-mini"):
         try:
@@ -87,6 +93,79 @@ class OpenAIClient:
         return repaired
 
 
+class HuggingFaceClient:
+    def __init__(self, api_token: str, model: str):
+        if not api_token:
+            raise ValueError("Hugging Face token required.")
+        try:
+            from huggingface_hub import InferenceClient  # type: ignore
+        except Exception as exc:  # pragma: no cover - import guard
+            raise RuntimeError(
+                "huggingface_hub package is required. Install with `pip install huggingface_hub`."
+            ) from exc
+
+        self.client = InferenceClient(model=model, token=api_token)
+        self.model = model
+
+    def chat(self, prompt: str, *, max_retries: int = 3) -> str:
+        delay = 1.0
+        last_error: Exception | None = None
+        for attempt in range(max_retries):
+            try:
+                return self.client.text_generation(
+                    prompt,
+                    max_new_tokens=1024,
+                    temperature=0.2,
+                    do_sample=False,
+                    return_full_text=False,
+                )
+            except Exception as exc:  # pragma: no cover - network call
+                last_error = exc
+                if _is_rate_limit_error(exc):
+                    wait_time = 30.0
+                    logger.warning(
+                        "Hugging Face rate limit encountered (attempt %s). Waiting %.1fs",
+                        attempt + 1,
+                        wait_time,
+                    )
+                    time.sleep(wait_time)
+                else:
+                    logger.warning(
+                        "Hugging Face call failed (attempt %s): %s", attempt + 1, exc
+                    )
+                    time.sleep(delay)
+                delay *= 2
+        raise RuntimeError(
+            f"Hugging Face call failed after retries: {last_error}"
+        )  # pragma: no cover - network call
+
+    def chat_json(self, prompt: str, *, max_retries: int = 3) -> Dict[str, Any]:
+        raw = self.chat(prompt, max_retries=max_retries)
+        parsed = _safe_json_parse(raw)
+        if parsed is not None:
+            return parsed
+
+        repair_prompt = (
+            "The previous response was invalid JSON. "
+            "Return ONLY valid JSON that fixes it without adding new facts.\n"
+            f"Original response:\n{raw}"
+        )
+        repaired_raw = self.chat(repair_prompt, max_retries=max_retries)
+        repaired = _safe_json_parse(repaired_raw)
+        if repaired is None:
+            raise ValueError("Model did not return valid JSON after repair attempt")
+        return repaired
+
+
+def build_client(provider: str, api_key: str, model: str) -> LLMClient:
+    normalized = provider.strip().lower()
+    if normalized in {"openai", "open ai"}:
+        return OpenAIClient(api_key=api_key, model=model)
+    if normalized in {"huggingface", "hugging face", "hugging face (inference api)"}:
+        return HuggingFaceClient(api_token=api_key, model=model)
+    raise ValueError(f"Unknown provider: {provider}")
+
+
 def _safe_json_parse(text: str) -> Dict[str, Any] | None:
     # Attempt direct parse
     try:

llm/pipeline.py

@@ -5,12 +5,14 @@ from typing import Tuple
 
 from schemas.resume import Resume, TailoredResume
 
-from .client import OpenAIClient
+from .client import build_client
 from .prompts import EXTRACTION_PROMPT, TAILORING_PROMPT
 
 
-def extract_resume_json(api_key: str, model: str, raw_text: str) -> Resume:
-    client = OpenAIClient(api_key=api_key, model=model)
+def extract_resume_json(
+    api_key: str, model: str, provider: str, raw_text: str
+) -> Resume:
+    client = build_client(provider, api_key=api_key, model=model)
     prompt = EXTRACTION_PROMPT.format(resume_text=raw_text)
     data = client.chat_json(prompt)
     resume = Resume.parse_obj(data)
@@ -21,12 +23,13 @@ def extract_resume_json(api_key: str, model: str, raw_text: str) -> Resume:
 def tailor_resume(
     api_key: str,
     model: str,
+    provider: str,
     resume: Resume,
     job_description: str,
     template_name: str,
     template_source: str,
 ) -> TailoredResume:
-    client = OpenAIClient(api_key=api_key, model=model)
+    client = build_client(provider, api_key=api_key, model=model)
     payload = json.loads(resume.json())
     prompt = TAILORING_PROMPT.format(
         resume_json=json.dumps(payload),
@@ -41,13 +44,20 @@ def tailor_resume(
 def run_pipeline(
     api_key: str,
     model: str,
+    provider: str,
     raw_text: str,
     job_description: str,
     template_name: str,
     template_source: str,
 ) -> Tuple[Resume, TailoredResume]:
-    resume = extract_resume_json(api_key, model, raw_text)
+    resume = extract_resume_json(api_key, model, provider, raw_text)
     tailored = tailor_resume(
-        api_key, model, resume, job_description, template_name, template_source
+        api_key,
+        model,
+        provider,
+        resume,
+        job_description,
+        template_name,
+        template_source,
     )
     return resume, tailored