improve prompts, templates, debugs, etc.

app.py

@@ -49,25 +49,26 @@ def run_refine(api_base, api_key, api_model, jd_info, cv_text):
     cheapAPI = {"base": api_base, "key": api_key, "model": api_model}
     taskAI = TaskAI(cheapAPI, temperature=0.2, max_tokens=2048)  # max_tokens=2048
     info("API initialized")
-    gen = (
+    gen = stream_together(
         taskAI.jd_preprocess(input=jd),
         taskAI.cv_preprocess(input=cv),
     )
-    info("tasks initialized")
-    result = [""] * 2
-    while 1:
-        stop: bool = True
-        for i in range(len(gen)):
-            try:
-                result[i] += next(gen[i]).delta
-                stop = False
-            except StopIteration:
-                # info(f"gen[{i}] exhausted")
-                pass
+    # result = [""] * 2
+    # while 1:
+    #     stop: bool = True
+    #     for i in range(len(gen)):
+    #         try:
+    #             result[i] += next(gen[i]).delta
+    #             stop = False
+    #         except StopIteration:
+    #             # info(f"gen[{i}] exhausted")
+    #             pass
+    #     yield result
+    #     if stop:
+    #         info("tasks done")
+    #         break
+    for result in gen:
         yield result
-        if stop:
-            info("tasks done")
-            break
 
 def run_compose(api_base, api_key, api_model, min_jd, min_cv):
     strongAPI = {"base": api_base, "key": api_key, "model": api_model}
@@ -78,28 +79,34 @@ def run_compose(api_base, api_key, api_model, min_jd, min_cv):
         result += response.delta
         yield result
 
-def finalize_letter_txt(api_base, api_key, api_model, debug_CoT, jd, cv):
+def finalize_letter_txt(api_base, api_key, api_model, debug_CoT):
     cheapAPI = {"base": api_base, "key": api_key, "model": api_model}
     taskAI = TaskAI(cheapAPI, temperature=0.2, max_tokens=2048)
     info("Finalizing letter ...")
-    gen = stream_together(
-        taskAI.purify_letter(full_text=debug_CoT),
-        taskAI.get_jobapp_meta(JD=jd, CV=cv),
-    )
-    for result in gen:
+    result=""
+    for response in taskAI.purify_letter(full_text=debug_CoT):
+        result += response.delta
         yield result
+    # gen = stream_together(
+    #     taskAI.purify_letter(full_text=debug_CoT),
+    # )
+    # for result in gen:
+    #     yield result
 
-def finalize_letter_pdf(meta_data, cover_letter_text):
+def finalize_letter_pdf(api_base, api_key, api_model, jd, cv, cover_letter_text):
+    cheapAPI = {"base": api_base, "key": api_key, "model": api_model}
+    taskAI = TaskAI(cheapAPI, temperature=0.1, max_tokens=100)
+    meta_data = next(taskAI.get_jobapp_meta(JD=jd, CV=cv))
     pdf_context = json.loads(meta_data)
     pdf_context["letter_body"] = cover_letter_text
-    return compile_pdf(pdf_context,tmpl_path="template_letter.tmpl",output_path=f"/tmp/cover_letter_{pdf_context['applicant_full_name']}_to_{pdf_context['company_full_name']}.pdf")
+    return meta_data, compile_pdf(pdf_context,tmpl_path="template_letter.tmpl",output_path=f"/tmp/cover_letter_by_{pdf_context['applicantFullName']}_to_{pdf_context['companyFullName']}.pdf")
 
 with gr.Blocks(
     title=DEMO_TITLE,
     theme=gr.themes.Base(primary_hue="blue", secondary_hue="sky", neutral_hue="slate"),
 ) as app:
     intro = f"""# {DEMO_TITLE}
-    > You provide job description and résumé. I write Cover letter for you!
+    > You provide job description and résumé. I write Cover letter for you!  
     Before you use, please fisrt setup API for 2 AI agents': Cheap AI and Strong AI.
     """
     gr.Markdown(intro)
@@ -114,7 +121,7 @@ with gr.Blocks(
                     cheap_base = gr.Textbox(
                         value=CHEAP_API_BASE, label="API BASE"
                     )
-                    cheap_key = gr.Textbox(value=CHEAP_API_KEY, label="API key")
+                    cheap_key = gr.Textbox(value=CHEAP_API_KEY, label="API key", type="password")
                     cheap_model = gr.Textbox(value=CHEAP_MODEL, label="Model ID")
                 gr.Markdown(
                     "---\n**Strong AI**, a thoughtful wordsmith, generates perfect cover letters to make both you and recruiters happy."
@@ -137,17 +144,17 @@ with gr.Blocks(
                 )
             with gr.Group():
                 gr.Markdown("## Applicant - CV / Résumé")
-                with gr.Row():
-                    cv_file = gr.File(
-                        label="Allowed formats: " + " ".join(CV_EXT),
-                        file_count="single",
-                        file_types=CV_EXT,
-                        type="filepath",
-                    )
-                    cv_text = gr.TextArea(
-                        label="Or enter text",
-                        placeholder="If attempting to both upload a file and enter text, only this text will be used.",
-                    )
+            # with gr.Row():
+                cv_file = gr.File(
+                    label="Allowed formats: " + " ".join(CV_EXT),
+                    file_count="single",
+                    file_types=CV_EXT,
+                    type="filepath",
+                )
+                cv_text = gr.TextArea(
+                    label="Or enter text",
+                    placeholder="If attempting to both upload a file and enter text, only this text will be used.",
+                )
         with gr.Column(scale=2):
             gr.Markdown("## Result")
             with gr.Accordion("Reformatting", open=True) as reformat_zone:
@@ -179,8 +186,8 @@ with gr.Blocks(
     ).then(fn=lambda:[gr.Accordion("Expert Zone", open=True),gr.Accordion("Reformatting", open=False)],inputs=None, outputs=[expert_zone, reformat_zone]
     ).then(fn=run_compose, inputs=[strong_base, strong_key, strong_model, min_jd, min_cv], outputs=[debug_CoT]                      
     ).then(fn=lambda:gr.Accordion("Expert Zone", open=False),inputs=None, outputs=[expert_zone]
-    ).then(fn=finalize_letter_txt, inputs=[cheap_base, cheap_key, cheap_model, debug_CoT, jd_info, cv_text], outputs=[cover_letter_text, debug_jobapp]
-    ).then(fn=finalize_letter_pdf, inputs=[debug_jobapp, cover_letter_text], outputs=[cover_letter_pdf])
+    ).then(fn=finalize_letter_txt, inputs=[cheap_base, cheap_key, cheap_model, debug_CoT], outputs=[cover_letter_text]
+    ).then(fn=finalize_letter_pdf, inputs=[cheap_base, cheap_key, cheap_model, jd_info, cv_text, cover_letter_text], outputs=[debug_jobapp, cover_letter_pdf])
 
 
 if __name__ == "__main__":

config.py

@@ -5,11 +5,11 @@ OPENAI_API_KEY = os.getenv("OPENAI_API_KEY") or ""
 
 CHEAP_API_BASE = os.getenv("CHEAP_API_BASE") or OPENAI_API_BASE
 CHEAP_API_KEY = os.getenv("CHEAP_API_KEY") or OPENAI_API_KEY
-CHEAP_MODEL = os.getenv("CHEAP_MODEL") or "gpt-4"
+CHEAP_MODEL = os.getenv("CHEAP_MODEL") or "gpt-3.5-turbo"
 
 STRONG_API_BASE = os.getenv("STRONG_API_BASE") or OPENAI_API_BASE
 STRONG_API_KEY = os.getenv("STRONG_API_KEY") or OPENAI_API_KEY
-STRONG_MODEL = os.getenv("STRONG_MODEL") or "mixtral-8x7b"
+STRONG_MODEL = os.getenv("STRONG_MODEL") or "gpt-4"
 
 IS_SHARE = bool(os.getenv("IS_SHARE")) or False
 

data_test.py

@@ -66,10 +66,10 @@ References
 Available upon request.
 """
 pdf_context = {
-  "company_full_name": "Mastercard",
-  "job_title": "Project Management Intern",
-  "applicant_full_name": "Dorothy Gale",
-  "applicant_contact_information": "123 Main St., Emerald City, KS 12345, (123) 456-7890, dorothy@wizardofoz.com",
+  "companyFullName": "Mastercard",
+  "jobTitle": "Project Management Intern",
+  "applicantFullName": "Dorothy Gale",
+  "applicantContactInformation": "123 Main St., Emerald City, KS 12345, (123) 456-7890, dorothy@wizardofoz.com",
   "letter_body": "text,\n\ntest test"
 }
 

taskAI.py

@@ -13,7 +13,7 @@ EXTRACT_INFO = ChatPromptTemplate(
     [
         ChatMessage(
             role="system",
-            content="You are a content extractor. You never paraphrase; you only reduce content at the sentence level. Your mission is to extract {to_extract} from user input. Make sure output is complete without missing parts. Output is in a clean text format",
+            content="You are a content extractor. You never paraphrase; you only reduce content at the sentence level. Your mission is to extract {to_extract} from user input. Reformat the extraction in a clean style if extraction looks messey.",
         ),
         ChatMessage(role="user", content="{input}"),
     ]
@@ -35,7 +35,7 @@ JSON_API = ChatPromptTemplate(
     [
         ChatMessage(
             role="system",
-            content="You are an AI JSON API. You convert user input into a JSON object. API returns exactly in this template: {template}",
+            content="You are a JSON API. Your mission is to convert user input into a JSON object exactly in this template: {template}",
         ),
         ChatMessage(role="user", content="{content}"),
     ]
@@ -51,7 +51,7 @@ LETTER_COMPOSE = ChatPromptTemplate(
 
 Before officially write the letter, think step by step. First, list what makes a perfect cover letter in general, and in order to write a perfect cover letter, what key points do you have to learn from the RESUME and JOB_DESCRIPTION. Then, carefully analyze the given RESUME and JOB_DESCRIPTION, take a deep breath and propose 3 best tactics to convince recruiter believe the applicant fit for the role. Ensure your thoughts are express clearly and then write the complete cover letter.""",
         ),
-        ChatMessage(role="user", content="<RESUME>\n{resume}\n</RESUME>\n\n<JOB_DESCRIPTION>\n{jd}</JOB_DESCRIPTION>\n"),
+        ChatMessage(role="user", content="<RESUME>\n{resume}\n</RESUME>\n\n<JOB_DESCRIPTION>\n{jd}</JOB_DESCRIPTION>\n<ANALYSIS_REPORT>"),
     ]
 )
 
@@ -82,7 +82,7 @@ class TaskAI(OpenAILike):
         )
 
     def jd_preprocess(self, input: str):
-        return self.stream_chat(EXTRACT_INFO.format_messages(to_extract="information directly related to job description", input=input))
+        return self.stream_chat(EXTRACT_INFO.format_messages(to_extract="the job description part`", input=input))
 
     def cv_preprocess(self, input: str):
         return self.stream_chat(SIMPLIFY_MD.format_messages(input=input))
@@ -91,9 +91,9 @@ class TaskAI(OpenAILike):
         return self.stream_chat(LETTER_COMPOSE.format_messages(resume=resume, jd=jd))
 
     def get_jobapp_meta(self, JD, CV):
-        meta_JD = self.chat(JSON_API.format_messages(template=keys_to_template(["company_full_name", "job_title"]), content=JD)).message.content
+        meta_JD = self.chat(JSON_API.format_messages(template=keys_to_template(["companyFullName", "jobTitle"]), content=JD)).message.content
         # yield meta_JD
-        meta_CV = self.chat(JSON_API.format_messages(template=keys_to_template(["applicant_full_name", "applicant_contact_information"]), content=CV)).message.content
+        meta_CV = self.chat(JSON_API.format_messages(template=keys_to_template(["applicantFullNname", "applicantContactInformation"]), content=CV)).message.content
         # yield meta_JD+'\n'+meta_CV
         try:
             meta_JD = json.loads(meta_JD.strip())
@@ -106,5 +106,5 @@ class TaskAI(OpenAILike):
         yield json.dumps(meta, indent=2)
 
     def purify_letter(self, full_text):
-        return self.stream_chat(EXTRACT_INFO.format_messages(to_extract="the letter part from greeting to sign-off, and remove applicant's name at end", input=full_text))
+        return self.stream_chat(EXTRACT_INFO.format_messages(to_extract="the cover letter section starting from 'Dear Hiring Manager' or similar to 'Sincerely,' or similar ", input=full_text))
 

taskNonAI.py

@@ -18,8 +18,10 @@ def extract_url(url: str) -> Optional[str]:
     cmd = f"""shot-scraper javascript -b firefox \
       "{url}" "
     async () => {{
+      const sleep = duration => new Promise(resolve => setTimeout(() => resolve(), duration));
       const readability = await import('https://cdn.skypack.dev/@mozilla/readability');
-      return (new readability.Readability(document)).parse();
+      await sleep(3000);
+      return new readability.Readability(document).parse();
     }}"
 """
     result = subprocess.run(cmd, shell=True, capture_output=True, text=True)
@@ -43,7 +45,7 @@ def date():
         f"%B %d{'th' if 4 <= current_date.day <= 20 or 24 <= current_date.day <= 30 else ['st', 'nd', 'rd'][current_date.day % 10 - 1]} , %Y")
 
 def typst_escape(s):
-    return s.replace('@','\@').replace('#','\#')
+    return str(s).replace('@','\@').replace('#','\#')
 
 def compile_pdf(context: dict, tmpl_path: str, output_path="/tmp/cover_letter.pdf"):
     with open(tmpl_path, "r", encoding='utf8') as f:

template_letter.tmpl

@@ -1,15 +1,15 @@
 #import "template_base.typ": *
 #show: letter.with(
   sender: [
-     ${applicant_contact_information}
+     ${applicantContactInformation}
   ],
   recipient: [
     Hiring Manager \
-    ${company_full_name} \
+    ${companyFullName} \
   ],
   date: [${date_string}],
-  subject: [Cover Letter for ${job_title}],
-  name: [${applicant_full_name}],
+  subject: [Cover Letter for ${jobTitle}],
+  name: [${applicantFullName}],
 )
 
 ${letter_body}