dhisana 0.0.1.dev272__py3-none-any.whl → 0.0.1.dev274__py3-none-any.whl

dhisana/utils/email_parse_helpers.py

@@ -1,19 +1,86 @@
 import base64
 import email.utils
+import quopri
+from email.message import Message
 from email.utils import parseaddr
 from typing import Any, Dict, List, Optional, Tuple
+
 from bs4 import BeautifulSoup
 
-def decode_base64_url(data: str) -> str:
+def decode_base64_url(data: str) -> bytes:
     """
-    Decodes a Base64-url-encoded string (Gmail API uses URL-safe Base64).
+    Decode a Base64-url-encoded string (Gmail API uses URL-safe Base64).
     """
     data = data.replace('-', '+').replace('_', '/')
     # Fix padding
     missing_padding = len(data) % 4
     if missing_padding:
         data += '=' * (4 - missing_padding)
-    return base64.b64decode(data).decode('utf-8', errors='ignore')
+    return base64.b64decode(data)
+
+
+def _get_charset(headers_list: List[Dict[str, str]], mime_type: str) -> Optional[str]:
+    content_type = find_header(headers_list, "Content-Type") or mime_type or ""
+    if not content_type:
+        return None
+    msg = Message()
+    msg["Content-Type"] = content_type
+    return msg.get_content_charset()
+
+
+def _decode_transfer_encoding(payload: bytes, transfer_encoding: str) -> bytes:
+    encoding = (transfer_encoding or "").lower()
+    if "quoted-printable" in encoding:
+        return quopri.decodestring(payload)
+    if "base64" in encoding:
+        try:
+            return base64.b64decode(payload, validate=False)
+        except Exception:
+            return payload
+    return payload
+
+
+_MOJIBAKE_MARKERS = (
+    "\u00e2\u0080\u0099",
+    "\u00e2\u0080\u0093",
+    "\u00e2\u0080\u0094",
+    "\u00e2\u0080\u009c",
+    "\u00e2\u0080\u009d",
+    "\u00e2\u0080\u00a6",
+    "\u00c3\u00a9",
+    "\u00c3\u00a0",
+    "\u00c3\u00b6",
+)
+
+
+def _repair_mojibake(text: str) -> str:
+    if not text or not any(marker in text for marker in _MOJIBAKE_MARKERS):
+        return text
+    try:
+        repaired = text.encode("latin-1").decode("utf-8")
+    except UnicodeError:
+        return text
+    if any(marker in repaired for marker in _MOJIBAKE_MARKERS):
+        return text
+    return repaired
+
+
+def _decode_part_text(
+    data: str, headers_list: List[Dict[str, str]], mime_type: str
+) -> str:
+    raw_bytes = decode_base64_url(data)
+    transfer_encoding = find_header(headers_list, "Content-Transfer-Encoding") or ""
+    decoded_bytes = _decode_transfer_encoding(raw_bytes, transfer_encoding)
+
+    charset = _get_charset(headers_list, mime_type)
+    tried = [charset, "utf-8", "windows-1252", "latin-1"]
+    for enc in [e for e in tried if e]:
+        try:
+            text = decoded_bytes.decode(enc)
+            return _repair_mojibake(text)
+        except (LookupError, UnicodeDecodeError):
+            continue
+    return _repair_mojibake(decoded_bytes.decode("utf-8", errors="replace"))
 
 
 def parse_plain_text_from_parts(parts: List[Dict[str, Any]]) -> str:
@@ -30,7 +97,9 @@ def parse_plain_text_from_parts(parts: List[Dict[str, Any]]) -> str:
             mime_type = part.get('mimeType', '')
             data = part.get('body', {}).get('data', '')
             if data:
-                decoded_data = decode_base64_url(data)
+                decoded_data = _decode_part_text(
+                    data, part.get("headers", []), mime_type
+                )
                 if 'text/plain' in mime_type:
                     text_chunks.append(decoded_data)
                 elif 'text/html' in mime_type:
@@ -49,7 +118,9 @@ def extract_email_body_in_plain_text(message_data: Dict[str, Any]) -> str:
     # If top-level body has data (i.e. single-part message)
     if payload.get('body', {}).get('data'):
         raw_data = payload['body']['data']
-        decoded_data = decode_base64_url(raw_data)
+        decoded_data = _decode_part_text(
+            raw_data, payload.get("headers", []), payload.get("mimeType", "")
+        )
         # Check if it might be HTML
         mime_type = payload.get('mimeType', '')
         if 'text/html' in mime_type:
@@ -85,7 +156,7 @@ def find_header(headers_list: List[Dict[str, str]], header_name: str) -> Optiona
     return None
 
 
-def parse_single_address(display_str: str) -> (str, str):
+def parse_single_address(display_str: str) -> Tuple[str, str]:
     """
     Parses a single display string like "Alice <alice@example.com>" 
     returning (name, email).
@@ -112,7 +183,7 @@ def parse_address_list(display_str: str) -> List[Tuple[str, str]]:
     return addresses
 
 
-def find_all_recipients_in_headers(headers_list: List[Dict[str, str]]) -> (str, str):
+def find_all_recipients_in_headers(headers_list: List[Dict[str, str]]) -> Tuple[str, str]:
     """
     Collect 'To', 'Cc', 'Bcc' headers, parse each address, and return:
       (comma-separated receiver names, comma-separated receiver emails)

dhisana/utils/generate_email.py

@@ -15,7 +15,25 @@ from pydantic import BaseModel, ConfigDict
 # ---------------------------------------------------------------------------------------
 # CONSTANTS
 # ---------------------------------------------------------------------------------------
-DEFAULT_OPENAI_MODEL = "gpt-4.1"
+DEFAULT_OPENAI_MODEL = "gpt-4.1"  # Larger context length - used for generating instructions
+DEFAULT_EMAIL_GEN_MODEL = "gpt-5.1-chat"  # Better content generation - used for generating email
+
+# -----------------------------------------------------------------------------
+# Email Generation Instructions schema (intermediate step)
+# -----------------------------------------------------------------------------
+class EmailGenerationInstructions(BaseModel):
+    """Instructions generated by the first step to guide email generation."""
+    recipient_context: str  # Key info about the lead/recipient
+    sender_context: str  # Key info about the sender
+    campaign_context: str  # Product, value prop, pain points, etc.
+    conversation_context: str  # Summary of any existing conversation
+    tone_and_style: str  # How the email should sound
+    key_points_to_include: str  # Main points to cover in the email
+    call_to_action: str  # What action to request
+    formatting_requirements: str  # HTML vs plain text, signature format, etc.
+    
+    model_config = ConfigDict(extra="forbid")
+
 
 # -----------------------------------------------------------------------------
 # Email Copy schema
@@ -68,28 +86,18 @@ FRAMEWORK_VARIATIONS = [
 ]
 
 # -----------------------------------------------------------------------------
-# Core function to generate an email copy
+# Step 1: Generate email instructions using gpt-4.1 (larger context)
 # -----------------------------------------------------------------------------
-async def generate_personalized_email_copy(
+async def generate_email_instructions(
     email_context: ContentGenerationContext,
     message_instructions: MessageGenerationInstructions,
     variation_text: str,
     tool_config: Optional[List[Dict]] = None,
-) -> dict:
+) -> EmailGenerationInstructions:
     """
-    Generate a personalized email using the provided context and instructions.
-
-    Steps:
-      1. Build a prompt referencing 6 main info:
-          (a) Lead Info
-          (b) Sender Info
-          (c) Campaign Info
-          (d) Messaging Instructions
-          (e) Additional Data (vector store) if any
-          (f) Current Conversation Context
-      2. Generate an initial draft with or without vector store usage.
-      3. Optionally refine if a vector store was used and user instructions were not provided.
-      4. Return the final subject & body.
+    Step 1 of 2-step email generation process.
+    Uses gpt-4.1 (larger context length) to process all context and generate
+    concise instructions for email generation.
     """
     cleaned_context = cleanup_email_context(email_context)
 
@@ -110,110 +118,212 @@ async def generate_personalized_email_copy(
     conversation_data = cleaned_context.current_conversation_context or ConversationContext()
 
     html_note = (
-        f"\n        Provide the HTML body using this guidance/template when possible:\n        {message_instructions.html_template}"
+        f"\n        HTML Template to follow: {message_instructions.html_template}"
         if getattr(message_instructions, "html_template", None)
         else ""
     )
-    important_requirements = """
-        IMPORTANT REQUIREMENTS:
+    
+    allow_html = getattr(message_instructions, "allow_html", False)
+    format_requirement = "HTML email with body_html field" if allow_html else "Plain text only (no HTML tags)"
+
+    # Construct the prompt for generating instructions
+    instructions_prompt = f"""
+Generate email writing instructions based on the following business context.
+
+LEAD INFORMATION:
+{lead_data.dict()}
+
+SENDER INFORMATION:
+Full Name: {sender_data.sender_full_name or ''}
+First Name: {sender_data.sender_first_name or ''}
+Last Name: {sender_data.sender_last_name or ''}
+Bio: {sender_data.sender_bio or ''}
+Appointment Booking URL: {sender_data.sender_appointment_booking_url or ''}
+
+CAMPAIGN INFORMATION:
+Product Name: {campaign_data.product_name or ''}
+Value Proposition: {campaign_data.value_prop or ''}
+Call To Action: {campaign_data.call_to_action or ''}
+Pain Points: {campaign_data.pain_points or []}
+Proof Points: {campaign_data.proof_points or []}
+Email Triage Guidelines: {campaign_data.email_triage_guidelines or ''}
+LinkedIn Triage Guidelines: {campaign_data.linkedin_triage_guidelines or ''}
+
+MESSAGING FRAMEWORK:
+{selected_instructions}{html_note}
+
+CONVERSATION HISTORY:
+Email Thread: {conversation_data.current_email_thread or ''}
+LinkedIn Thread: {conversation_data.current_linkedin_thread or ''}
+
+OUTPUT FIELDS:
+- recipient_context: Key details about the lead (name, company, role, relevant background)
+- sender_context: Sender name and relevant info for signature
+- campaign_context: Product or service being promoted, value proposition, key pain points
+- conversation_context: Summary of any prior conversation if exists
+- tone_and_style: Professional, friendly, etc.
+- key_points_to_include: Specific points to mention in the email
+- call_to_action: What action the recipient should take
+- formatting_requirements: Format is {format_requirement}. Salutation format, signature requirements.
+    """
+
+    # Check if a vector store is available
+    vector_store_id = (email_context.external_known_data.external_openai_vector_store_id
+                       if email_context.external_known_data else None)
+
+    instructions_response = None
+    instructions_status = ""
+
+    # Generate instructions using gpt-4.1 (larger context)
+    if vector_store_id:
+        instructions_response, instructions_status = await get_structured_output_with_assistant_and_vector_store(
+            prompt=instructions_prompt,
+            response_format=EmailGenerationInstructions,
+            vector_store_id=vector_store_id,
+            model=DEFAULT_OPENAI_MODEL,
+            tool_config=tool_config,
+            use_cache=email_context.message_instructions.use_cache if email_context.message_instructions else True
+        )
+    else:
+        instructions_response, instructions_status = await get_structured_output_internal(
+            prompt=instructions_prompt,
+            response_format=EmailGenerationInstructions,
+            model=DEFAULT_OPENAI_MODEL,
+            tool_config=tool_config,
+            use_cache=email_context.message_instructions.use_cache if email_context.message_instructions else True
+        )
+
+    if instructions_status != "SUCCESS":
+        raise Exception("Error: Could not generate email instructions in step 1.")
+    
+    return instructions_response
+
+
+# -----------------------------------------------------------------------------
+# Step 2: Generate email using gpt-5.1-chat (better content generation)
+# -----------------------------------------------------------------------------
+async def generate_email_from_instructions(
+    instructions: EmailGenerationInstructions,
+    message_instructions: MessageGenerationInstructions,
+    tool_config: Optional[List[Dict]] = None,
+    use_cache: bool = True,
+) -> EmailCopy:
+    """
+    Step 2 of 2-step email generation process.
+    Uses gpt-5.1-chat (better content generation) to generate the email
+    from the condensed instructions.
+    """
+    allow_html = getattr(message_instructions, "allow_html", False)
+    
+    if allow_html:
+        output_requirements = """
+        OUTPUT REQUIREMENTS:
         - Output must be JSON with "subject", "body", and "body_html" fields.
         - "body_html" should be clean HTML suitable for email (no external assets), inline styles welcome.
         - "body" must be the plain-text equivalent of "body_html".
-        - Keep it concise and relevant. No placeholders or extra instructions.
-        - Do not include PII or internal references, guids or content identifiers in the email.
-        - Use conversational names for company/person placeholders when provided.
-        - Email has salutation Hi <First Name>, unless otherwise specified.
-        - Make sure the signature in body has the sender_first_name correct and in the format the user specified.
-        - Do Not Make up information. use the information provided in the context and instructions only.
-        - Do Not use em dash in the generated output.
-    """
-    if not getattr(message_instructions, "allow_html", False):
-        important_requirements = """
-        IMPORTANT REQUIREMENTS:
+        """
+    else:
+        output_requirements = """
+        OUTPUT REQUIREMENTS:
         - Output must be JSON with "subject" and "body" fields only.
         - In the subject or body DO NOT include any HTML tags like <a>, <b>, <i>, etc.
         - The body and subject should be in plain text.
-        - If there is a link provided in email use it as is. dont wrap it in any HTML tags.
-        - Keep it concise and relevant. No placeholders or extra instructions.
-        - Do not include PII or internal references, guids or content identifiers in the email.
-        - User conversational name for company name if used.
-        - Email has saluation Hi <First Name>, unless otherwise specified.
-        - <First Name> is the first name of the lead. Its conversational name. It does not have any special characters or spaces.
-        - Make sure the signature in body has the sender_first_name is correct and in the format user has specified.
-        - Do Not Make up information. use the information provided in the context and instructions only.
-        - Make sure the body text is well-formatted and that newline and carriage-return characters are correctly present and preserved in the message body.
-        - Do Not use em dash in the generated output.
-    """
+        - If there is a link provided, use it as is without wrapping in any HTML tags.
+        """
+
+    email_prompt = f"""
+Generate a personalized business email based on these specifications:
 
-    # Construct the consolidated prompt
-    initial_prompt = f"""
-        Hi AI Assistant,
+RECIPIENT CONTEXT:
+{instructions.recipient_context}
 
-        Below is the context in 6 main sections. Use it to craft a concise, professional email:
+SENDER CONTEXT:
+{instructions.sender_context}
 
-        1) Lead Information:
-        {lead_data.dict()}
+CAMPAIGN CONTEXT:
+{instructions.campaign_context}
 
-        2) Sender Information:
-        Full Name: {sender_data.sender_full_name or ''}
-        First Name: {sender_data.sender_first_name or ''}
-        Last Name: {sender_data.sender_last_name or ''}
-        Bio: {sender_data.sender_bio or ''}
-        Appointment Booking URL: {sender_data.sender_appointment_booking_url or ''}
+CONVERSATION CONTEXT:
+{instructions.conversation_context}
 
-        3) Campaign Information:
-        Product Name: {campaign_data.product_name or ''}
-        Value Proposition: {campaign_data.value_prop or ''}
-        Call To Action: {campaign_data.call_to_action or ''}
-        Pain Points: {campaign_data.pain_points or []}
-        Proof Points: {campaign_data.proof_points or []}
-        Triage Guidelines (Email): {campaign_data.email_triage_guidelines or ''}
-        Triage Guidelines (LinkedIn): {campaign_data.linkedin_triage_guidelines or ''}
+TONE AND STYLE:
+{instructions.tone_and_style}
 
-        4) Messaging Instructions (template/framework):
-        {selected_instructions}{html_note}
+KEY POINTS TO INCLUDE:
+{instructions.key_points_to_include}
 
-        5) External Data / Vector Store:
-        (I will be provided with file_search tool if present.)
+CALL TO ACTION:
+{instructions.call_to_action}
 
-        6) Current Conversation Context:
-        Email Thread: {conversation_data.current_email_thread or ''}
-        LinkedIn Thread: {conversation_data.current_linkedin_thread or ''}
+FORMATTING REQUIREMENTS:
+{instructions.formatting_requirements}
 
-        {important_requirements}
+{output_requirements}
+
+ADDITIONAL REQUIREMENTS:
+- Keep it concise and relevant. No placeholders.
+- Do not include internal references or content identifiers in the email.
+- Use only the information provided.
+- Ensure the body text is well-formatted with proper newlines.
+- Do not use em dash in the output.
     """
 
-    # Check if a vector store is available
-    vector_store_id = (email_context.external_known_data.external_openai_vector_store_id
-                       if email_context.external_known_data else None)
+    # Generate email using gpt-5.1-chat (better content generation)
+    email_response, email_status = await get_structured_output_internal(
+        prompt=email_prompt,
+        response_format=EmailCopy,
+        model=DEFAULT_EMAIL_GEN_MODEL,
+        tool_config=tool_config,
+        use_cache=use_cache
+    )
 
-    initial_response = None
-    initial_status = ""
+    if email_status != "SUCCESS":
+        raise Exception("Error: Could not generate email in step 2.")
+    
+    return email_response
 
-    # Generate initial draft
-    if vector_store_id:
-        initial_response, initial_status = await get_structured_output_with_assistant_and_vector_store(
-            prompt=initial_prompt,
-            response_format=EmailCopy,
-            vector_store_id=vector_store_id,
-            model=DEFAULT_OPENAI_MODEL,
-            tool_config=tool_config,
-            use_cache=email_context.message_instructions.use_cache if email_context.message_instructions else True
-        )
-    else:
-        # Otherwise, generate the initial draft internally
-        initial_response, initial_status = await get_structured_output_internal(
-            prompt=initial_prompt,
-            response_format=EmailCopy,
-            model=DEFAULT_OPENAI_MODEL,
-            tool_config=tool_config,
-            use_cache=email_context.message_instructions.use_cache if email_context.message_instructions else True
-        )
 
-    if initial_status != "SUCCESS":
-        raise Exception("Error: Could not generate initial draft for the personalized email.")
-    plain_body = initial_response.body
-    html_body = getattr(initial_response, "body_html", None)
+# -----------------------------------------------------------------------------
+# Core function to generate an email copy (2-step process)
+# -----------------------------------------------------------------------------
+async def generate_personalized_email_copy(
+    email_context: ContentGenerationContext,
+    message_instructions: MessageGenerationInstructions,
+    variation_text: str,
+    tool_config: Optional[List[Dict]] = None,
+) -> dict:
+    """
+    Generate a personalized email using a 2-step process:
+
+    Step 1: Use gpt-4.1 (larger context length) to process all context and generate
+            concise instructions for email generation.
+    
+    Step 2: Use gpt-5.1-chat (better content generation) to generate the actual
+            email from the condensed instructions.
+
+    This approach leverages gpt-4.1's larger context window to process extensive
+    lead/campaign data, and gpt-5.1-chat's superior content generation for the
+    final email output.
+    """
+    # Step 1: Generate instructions using gpt-4.1
+    instructions = await generate_email_instructions(
+        email_context=email_context,
+        message_instructions=message_instructions,
+        variation_text=variation_text,
+        tool_config=tool_config,
+    )
+
+    # Step 2: Generate email using gpt-5.1-chat
+    use_cache = email_context.message_instructions.use_cache if email_context.message_instructions else True
+    email_response = await generate_email_from_instructions(
+        instructions=instructions,
+        message_instructions=message_instructions,
+        tool_config=tool_config,
+        use_cache=use_cache,
+    )
+
+    plain_body = email_response.body
+    html_body = getattr(email_response, "body_html", None)
     if not plain_body and html_body:
         plain_body = _html_to_plain_text(html_body)
 
@@ -225,7 +335,7 @@ async def generate_personalized_email_copy(
         receiver_name=email_context.lead_info.full_name or "",
         receiver_email=email_context.lead_info.email or "",
         iso_datetime=datetime.utcnow().isoformat(),
-        subject=initial_response.subject,
+        subject=email_response.subject,
         body=plain_body,
         html_body=html_body if getattr(message_instructions, "allow_html", False) else None,
     )

dhisana/utils/generate_email_response.py

@@ -71,7 +71,6 @@ async def get_inbound_email_triage_action(
         "SCHEDULE_MEETING",
         "SEND_REPLY",
         "OOF_MESSAGE",
-        "NEED_MORE_INFO",
         "FORWARD_TO_OTHER_USER",
         "NO_MORE_IN_ORGANIZATION",
         "OBJECTION_RAISED",
@@ -93,6 +92,8 @@ async def get_inbound_email_triage_action(
         allowed_actions = 
         {allowed_actions}
 
+        If you need more info, use SEND_REPLY and ask a short clarifying question.
+
         1. Email thread or conversation:
         {[thread_item.model_dump() for thread_item in cleaned_context.current_conversation_context.current_email_thread]}
 
@@ -129,8 +130,7 @@ async def get_inbound_email_triage_action(
         Handling interest & objections
         ------------------------------
         • If the prospect asks for **pricing, docs, case studies, or more info**  
-        → **NEED_MORE_INFO** and craft a short response that promises to send
-            the material (or includes it if ≤ 150 words fits).
+        → **SEND_REPLY** and ask a short clarifying question or promise a follow-up.
 
         • If they mention **budget, timing, or competitor concerns**  
         → **OBJECTION_RAISED** and reply with a brief acknowledgement
@@ -183,7 +183,7 @@ async def get_inbound_email_triage_action(
         "triage_status": "...",
         "triage_reason": null or "<reason>",
         "response_action_to_take": "one of {allowed_actions}",
-        "response_message": "<only if SEND_REPLY/SCHEDULE_MEETING, else empty>"
+        "response_message": "<only if SEND_REPLY/SCHEDULE_MEETING/OBJECTION_RAISED, else empty>"
         }}
 
         Current date is: {current_date_iso}.
@@ -251,7 +251,6 @@ async def generate_inbound_email_response_copy(
         "UNSUBSCRIBE",
         "OOF_MESSAGE",
         "NOT_INTERESTED",
-        "NEED_MORE_INFO",
         "FORWARD_TO_OTHER_USER",
         "NO_MORE_IN_ORGANIZATION",
         "OBJECTION_RAISED",
@@ -282,7 +281,7 @@ async def generate_inbound_email_response_copy(
         if cleaned_context.current_conversation_context.current_email_thread else []}
 
     2) Lead information:
-    {lead_data.dict()}
+    {lead_data.model_dump()}
 
     Sender information:
     - Full name: {sender_data.sender_full_name or ''}
@@ -316,16 +315,17 @@ async def generate_inbound_email_response_copy(
     Choose exactly ONE action from:
     {allowed_actions}
 
+    If you need more info, use SEND_REPLY and ask a short clarifying question.
+
     Priority order:
     1. UNSUBSCRIBE
     2. NOT_INTERESTED
     3. OOF_MESSAGE
     4. SCHEDULE_MEETING
     5. FORWARD_TO_OTHER_USER
-    6. NEED_MORE_INFO
-    7. OBJECTION_RAISED
-    8. SEND_REPLY
-    9. END_CONVERSATION
+    6. OBJECTION_RAISED
+    7. SEND_REPLY
+    8. END_CONVERSATION
 
     =====================================================
     HOW THE RESPONSE SHOULD SOUND
@@ -391,7 +391,7 @@ async def generate_inbound_email_response_copy(
     "triage_status": "AUTOMATIC" or "END_CONVERSATION",
     "triage_reason": "<string if END_CONVERSATION, otherwise null>",
     "response_action_to_take": "one of {allowed_actions}",
-    "response_message": "<reply body only if SEND_REPLY or SCHEDULE_MEETING, otherwise empty>"
+    "response_message": "<reply body only if SEND_REPLY/SCHEDULE_MEETING/OBJECTION_RAISED, otherwise empty>"
     }}
 
     =====================================================
@@ -446,6 +446,12 @@ async def generate_inbound_email_response_copy(
                 f"Campaign ID: {campaign_id}"
             )
 
+    response_action = initial_response.response_action_to_take
+    if response_action == "NEED_MORE_INFO":
+        response_action = "SEND_REPLY"
+
+    response_message = initial_response.response_message or ""
+
     response_item = MessageItem(
         message_id="",  # or generate one if appropriate
         thread_id="",
@@ -455,7 +461,7 @@ async def generate_inbound_email_response_copy(
         receiver_email=campaign_context.lead_info.email or "",
         iso_datetime=datetime.datetime.utcnow().isoformat(),
         subject="",  # or set some triage subject if needed
-        body=initial_response.response_message
+        body=response_message
     )
 
     # Build a MessageResponse that includes triage metadata plus your message item
@@ -463,9 +469,8 @@ async def generate_inbound_email_response_copy(
         triage_status=initial_response.triage_status,
         triage_reason=initial_response.triage_reason,
         message_item=response_item,
-        response_action_to_take=initial_response.response_action_to_take
+        response_action_to_take=response_action
     )
-    print(response_message.model_dump())
     return response_message.model_dump()
 
 

dhisana/utils/generate_linkedin_connect_message.py

@@ -24,7 +24,22 @@ from dhisana.utils.assistant_tool_tag import assistant_tool
 # ---------------------------------------------------------------------------------------
 # CONSTANTS
 # ---------------------------------------------------------------------------------------
-DEFAULT_OPENAI_MODEL = "gpt-4.1"
+DEFAULT_OPENAI_MODEL = "gpt-4.1"  # Larger context length - used for generating instructions
+DEFAULT_EMAIL_GEN_MODEL = "gpt-5.1-chat"  # Better content generation - used for generating message
+
+
+# ----------------------------------------------------------------------
+# LinkedIn Generation Instructions Schema (intermediate step)
+# ----------------------------------------------------------------------
+class LinkedInGenerationInstructions(BaseModel):
+    """Instructions generated by the first step to guide LinkedIn message generation."""
+    recipient_context: str  # Key info about the lead/recipient
+    sender_context: str  # Key info about the sender
+    campaign_context: str  # Product, value prop, pain points, etc.
+    conversation_context: str  # Summary of any existing conversation
+    tone_and_style: str  # How the message should sound
+    key_points_to_include: str  # Main points to cover
+    call_to_action: str  # What action to request
 
 
 # ----------------------------------------------------------------------
@@ -68,26 +83,17 @@ LINKEDIN_FRAMEWORK_VARIATIONS = [
 ]
 
 # ----------------------------------------------------------------------
-# Core function to generate a LinkedIn copy (similar to generate_personalized_email_copy)
+# Step 1: Generate LinkedIn instructions using gpt-4.1 (larger context)
 # ----------------------------------------------------------------------
-async def generate_personalized_linkedin_copy(
+async def generate_linkedin_instructions(
     linkedin_context: ContentGenerationContext,
     variation_text: str,
     tool_config: Optional[List[Dict]] = None,
-) -> dict:
+) -> LinkedInGenerationInstructions:
     """
-    Generate a personalized LinkedIn connection message using the provided context and instructions.
-
-    Steps:
-      1. Build a prompt referencing 6 main sections:
-          (a) Lead Info
-          (b) Sender Info
-          (c) Campaign Info
-          (d) Variation / Instructions
-          (e) External Data (if any)
-          (f) Current Conversation
-      2. Generate the LinkedIn message with or without vector store usage.
-      3. Return the final subject & body, ensuring < 40 words.
+    Step 1 of 2-step LinkedIn message generation process.
+    Uses gpt-4.1 (larger context length) to process all context and generate
+    concise instructions for message generation.
     """
     cleaned_context = cleanup_linkedin_context(linkedin_context)
 
@@ -96,54 +102,43 @@ async def generate_personalized_linkedin_copy(
     campaign_data = cleaned_context.campaign_context or CampaignContext()
     conversation_data = cleaned_context.current_conversation_context or ConversationContext()
 
-    # Construct the consolidated prompt
-    prompt = f"""
-        Hi AI Assistant,
-
-        Below is the context in 6 main sections. Use it to craft a concise, professional LinkedIn connection request message:
-        Linked in connect message HAS to be maximum of 40 words. DO NOT go more than 40 words.
-
-        1) Lead Information:
-        {lead_data.dict()}
-
-        2) Sender Information:
-        Full Name: {sender_data.sender_full_name or ''}
-        First Name: {sender_data.sender_first_name or ''}
-        Last Name: {sender_data.sender_last_name or ''}
-        Bio: {sender_data.sender_bio or ''}
-        Appointment Booking URL: {sender_data.sender_appointment_booking_url or ''}
-
-        3) Campaign Information:
-        Product Name: {campaign_data.product_name or ''}
-        Value Proposition: {campaign_data.value_prop or ''}
-        Call To Action: {campaign_data.call_to_action or ''}
-        Pain Points: {campaign_data.pain_points or []}
-        Proof Points: {campaign_data.proof_points or []}
-
-        4) Follow these Instructions to generate the likedin connection request message:
-        {variation_text}
-
-        5) External Data / Vector Store:
-        (File search or additional context if available.)
-
-        6) Current Conversation Context:
-        Email Thread: {conversation_data.current_email_thread or ''}
-        LinkedIn Thread: {conversation_data.current_linkedin_thread or ''}
-
-        IMPORTANT REQUIREMENTS:
-        - Output must be JSON "body" of the message.
-        - The entire message body must be under 40 words total.
-        - In the body DO NOT include any HTML tags like <a>, <b>, <i>, etc.
-        - The body should be in plain text.
-        - If there is a link provided in body use it as is. dont wrap it in any HTML tags.
-        - No placeholders or extra instructions in the final output.
-        - Do not include personal addresses, IDs, or irrelevant internal data.
-        - Linked in message always has saluatation Hi <First Name>, unless specified otherwise.
-        - <First Name> is the first name of the lead. Its conversational name. It does not have any special characters or spaces.
-        - Make sure the signature in body has the sender_first_name is correct and in the format user has specified.
-        - User conversational name for company name if used.
-        - Make sure the body text is well-formatted and that newline and carriage-return characters are correctly present and preserved in the message body.
-        - Do Not use em dash in the generated output.
+    # Construct the prompt for generating instructions
+    instructions_prompt = f"""
+Generate LinkedIn connection message instructions based on the following business context.
+The final message must be under 40 words.
+
+LEAD INFORMATION:
+{lead_data.dict()}
+
+SENDER INFORMATION:
+Full Name: {sender_data.sender_full_name or ''}
+First Name: {sender_data.sender_first_name or ''}
+Last Name: {sender_data.sender_last_name or ''}
+Bio: {sender_data.sender_bio or ''}
+Appointment Booking URL: {sender_data.sender_appointment_booking_url or ''}
+
+CAMPAIGN INFORMATION:
+Product Name: {campaign_data.product_name or ''}
+Value Proposition: {campaign_data.value_prop or ''}
+Call To Action: {campaign_data.call_to_action or ''}
+Pain Points: {campaign_data.pain_points or []}
+Proof Points: {campaign_data.proof_points or []}
+
+MESSAGING FRAMEWORK:
+{variation_text}
+
+CONVERSATION HISTORY:
+Email Thread: {conversation_data.current_email_thread or ''}
+LinkedIn Thread: {conversation_data.current_linkedin_thread or ''}
+
+OUTPUT FIELDS:
+- recipient_context: Key details about the lead (name, company, role)
+- sender_context: Sender name and relevant info for signature
+- campaign_context: Product or service, value proposition, key pain points
+- conversation_context: Summary of any prior conversation if exists
+- tone_and_style: Professional, friendly, etc.
+- key_points_to_include: Main points for the message (keep brief for 40 word limit)
+- call_to_action: What action the recipient should take
     """
 
     vector_store_id = (
@@ -151,28 +146,131 @@ async def generate_personalized_linkedin_copy(
         if linkedin_context.external_known_data else None
     )
 
+    instructions_response = None
+    instructions_status = ""
+
+    # Generate instructions using gpt-4.1 (larger context)
     if vector_store_id:
-        # Use the vector store if available
-        response_data, status = await get_structured_output_with_assistant_and_vector_store(
-            prompt=prompt,
-            response_format=LinkedInConnectMessage,
+        instructions_response, instructions_status = await get_structured_output_with_assistant_and_vector_store(
+            prompt=instructions_prompt,
+            response_format=LinkedInGenerationInstructions,
             vector_store_id=vector_store_id,
             model=DEFAULT_OPENAI_MODEL,
             tool_config=tool_config,
             use_cache=linkedin_context.message_instructions.use_cache if linkedin_context.message_instructions else True
         )
     else:
-        # Otherwise, generate internally
-        response_data, status = await get_structured_output_internal(
-            prompt=prompt,
-            response_format=LinkedInConnectMessage,
+        instructions_response, instructions_status = await get_structured_output_internal(
+            prompt=instructions_prompt,
+            response_format=LinkedInGenerationInstructions,
             model=DEFAULT_OPENAI_MODEL,
             tool_config=tool_config,
             use_cache=linkedin_context.message_instructions.use_cache if linkedin_context.message_instructions else True
         )
 
-    if status != "SUCCESS":
-        raise Exception("Error: Could not generate the LinkedIn message.")
+    if instructions_status != "SUCCESS":
+        raise Exception("Error: Could not generate LinkedIn instructions in step 1.")
+    
+    return instructions_response
+
+
+# ----------------------------------------------------------------------
+# Step 2: Generate LinkedIn message using gpt-5.1-chat (better content)
+# ----------------------------------------------------------------------
+async def generate_linkedin_from_instructions(
+    instructions: LinkedInGenerationInstructions,
+    tool_config: Optional[List[Dict]] = None,
+    use_cache: bool = True,
+) -> LinkedInConnectMessage:
+    """
+    Step 2 of 2-step LinkedIn message generation process.
+    Uses gpt-5.1-chat (better content generation) to generate the message
+    from the condensed instructions.
+    """
+    message_prompt = f"""
+Generate a personalized LinkedIn connection request message based on these specifications.
+The message must be under 40 words total.
+
+RECIPIENT CONTEXT:
+{instructions.recipient_context}
+
+SENDER CONTEXT:
+{instructions.sender_context}
+
+CAMPAIGN CONTEXT:
+{instructions.campaign_context}
+
+CONVERSATION CONTEXT:
+{instructions.conversation_context}
+
+TONE AND STYLE:
+{instructions.tone_and_style}
+
+KEY POINTS TO INCLUDE:
+{instructions.key_points_to_include}
+
+CALL TO ACTION:
+{instructions.call_to_action}
+
+OUTPUT REQUIREMENTS:
+- Output must be JSON with "body" field only.
+- The entire message body must be under 40 words total.
+- Plain text only, no HTML tags.
+- Use salutation Hi followed by first name, unless specified otherwise.
+- Use only the information provided.
+- Do not use em dash in the output.
+    """
+
+    # Generate message using gpt-5.1-chat (better content generation)
+    message_response, message_status = await get_structured_output_internal(
+        prompt=message_prompt,
+        response_format=LinkedInConnectMessage,
+        model=DEFAULT_EMAIL_GEN_MODEL,
+        tool_config=tool_config,
+        use_cache=use_cache
+    )
+
+    if message_status != "SUCCESS":
+        raise Exception("Error: Could not generate LinkedIn message in step 2.")
+    
+    return message_response
+
+
+# ----------------------------------------------------------------------
+# Core function to generate a LinkedIn copy (2-step process)
+# ----------------------------------------------------------------------
+async def generate_personalized_linkedin_copy(
+    linkedin_context: ContentGenerationContext,
+    variation_text: str,
+    tool_config: Optional[List[Dict]] = None,
+) -> dict:
+    """
+    Generate a personalized LinkedIn connection message using a 2-step process:
+
+    Step 1: Use gpt-4.1 (larger context length) to process all context and generate
+            concise instructions for message generation.
+    
+    Step 2: Use gpt-5.1-chat (better content generation) to generate the actual
+            LinkedIn message from the condensed instructions.
+    """
+    cleaned_context = cleanup_linkedin_context(linkedin_context)
+    lead_data = cleaned_context.lead_info or Lead()
+    sender_data = cleaned_context.sender_info or SenderInfo()
+
+    # Step 1: Generate instructions using gpt-4.1
+    instructions = await generate_linkedin_instructions(
+        linkedin_context=linkedin_context,
+        variation_text=variation_text,
+        tool_config=tool_config,
+    )
+
+    # Step 2: Generate message using gpt-5.1-chat
+    use_cache = linkedin_context.message_instructions.use_cache if linkedin_context.message_instructions else True
+    message_response = await generate_linkedin_from_instructions(
+        instructions=instructions,
+        tool_config=tool_config,
+        use_cache=use_cache,
+    )
 
     # Wrap in MessageItem for consistency
     response_item = MessageItem(
@@ -184,7 +282,7 @@ async def generate_personalized_linkedin_copy(
         receiver_email=lead_data.email or "",
         iso_datetime=datetime.utcnow().isoformat(),
         subject="Hi",
-        body=response_data.body
+        body=message_response.body
     )
 
     return response_item.model_dump()

dhisana/utils/openai_assistant_and_file_utils.py

@@ -9,7 +9,7 @@ import json
 import logging
 import re
 import traceback
-from typing import Any, Dict, List, Optional
+from typing import Any, Dict, List, Optional, Tuple
 
 from fastapi import HTTPException
 
@@ -52,6 +52,14 @@ async def delete_vector_store(
     client = create_openai_client(tool_config)
     try:
         client.vector_stores.delete(vector_store_id=vector_store_id)
+    except openai.NotFoundError:
+        logging.warning(f"Vector store not found during delete: {vector_store_id}")
+    except openai.APIStatusError as e:
+        if getattr(e, "status_code", None) == 404:
+            logging.warning(f"Vector store not found during delete: {vector_store_id}")
+            return
+        logging.error(f"Error deleting vector store {vector_store_id}: {e}")
+        raise HTTPException(status_code=400, detail=str(e))
     except Exception as e:
         logging.error(f"Error deleting vector store {vector_store_id}: {e}")
         raise HTTPException(status_code=400, detail=str(e))
@@ -74,10 +82,11 @@ async def upload_file_openai_and_vector_store(
 
     try:
         if isinstance(file_path_or_bytes, str):
-            file_upload = client.files.create(
-                file=open(file_path_or_bytes, "rb"),
-                purpose=purpose,
-            )
+            with open(file_path_or_bytes, "rb") as f:
+                file_upload = client.files.create(
+                    file=f,
+                    purpose=purpose,
+                )
         elif isinstance(file_path_or_bytes, bytes):
             file_upload = client.files.create(
                 file=(file_name, file_path_or_bytes, mime_type),
@@ -108,10 +117,11 @@ async def upload_file_openai(
 
     try:
         if isinstance(file_path_or_bytes, str):
-            file_upload = client.files.create(
-                file=open(file_path_or_bytes, "rb"),
-                purpose=purpose,
-            )
+            with open(file_path_or_bytes, "rb") as f:
+                file_upload = client.files.create(
+                    file=f,
+                    purpose=purpose,
+                )
         else:
             file_upload = client.files.create(
                 file=(file_name, file_path_or_bytes, mime_type),
@@ -217,7 +227,7 @@ async def run_response_text(
     max_tokens: int = 2048,
     store: bool = True,
     tool_config: Optional[List[Dict]] = None,
-) -> (str, str):
+) -> Tuple[str, str]:
     """Plain text completion via the Responses API."""
     client = create_openai_client(tool_config)
 
@@ -241,7 +251,7 @@ async def run_response_structured(
     max_tokens: int = 1024,
     store: bool = True,
     tool_config: Optional[List[Dict]] = None,
-) -> (Any, str):
+) -> Tuple[Any, str]:
     """Structured JSON output via Responses API."""
     client = create_openai_client(tool_config)
 

{dhisana-0.0.1.dev272.dist-info → dhisana-0.0.1.dev274.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: dhisana
-Version: 0.0.1.dev272
+Version: 0.0.1.dev274
 Summary: A Python SDK for Dhisana AI Platform
 Home-page: https://github.com/dhisana-ai/dhisana-python-sdk
 Author: Admin

{dhisana-0.0.1.dev272.dist-info → dhisana-0.0.1.dev274.dist-info}/RECORD

@@ -30,7 +30,7 @@ dhisana/utils/composite_tools.py,sha256=ZlwHCp7PXjYFUWUEeR_fTF0Z4Wg-4F6eBi1reE3F
 dhisana/utils/dataframe_tools.py,sha256=R6eUXjwR5SG6_K87rWjj4T5PT2w6xvVF2EKBajIv-RE,9242
 dhisana/utils/domain_parser.py,sha256=Kw5MPP06wK2azWQzuSiOE-DffOezLqDyF-L9JEBsMSU,1206
 dhisana/utils/email_body_utils.py,sha256=rlCVjdBlqNnEiUberJGXGcrYY1GQOkW0-aB6AEpS3L4,2302
-dhisana/utils/email_parse_helpers.py,sha256=LIdm1B1IyGSW50y8EkxOk6YRjvxO2SJTgTKPLxYls_o,4613
+dhisana/utils/email_parse_helpers.py,sha256=rl72ggS-yoB-w3ZHW2sevKJulQ-_8iLdpVTH6QnKPcs,6789
 dhisana/utils/email_provider.py,sha256=ukW_0nHcjTQmpnE9pdJci78LrZcsK1_0v6kcgc2ChPY,14573
 dhisana/utils/enrich_lead_information.py,sha256=O0fV-8MlXFT_z5aXvmvXVT76AISN94GpvAOlq3q_Phw,39411
 dhisana/utils/extract_email_content_for_llm.py,sha256=SQmMZ3YJtm3ZI44XiWEVAItcAwrsSSy1QzDne7LTu_Q,3713
@@ -39,11 +39,11 @@ dhisana/utils/field_validators.py,sha256=BZgNCpBG264aRqNUu_J67c6zfr15zlAaIw2XRy8
 dhisana/utils/g2_tools.py,sha256=a4vmBYCBvLae5CdpOhMN1oNlvO8v9J1B5Sd8T5PzuU8,3346
 dhisana/utils/generate_content.py,sha256=kkf-aPuA32BNgwk_j5N6unYHOZpO7zIfO6zP95XM9fA,2298
 dhisana/utils/generate_custom_message.py,sha256=tQsryytoYKP5uF3bRENeZks1LvOMFCP6L1487P_r_hk,12072
-dhisana/utils/generate_email.py,sha256=Qufol-qgOk8loNulZ6hMzIRr45Ml7JwJpLO5OMabbH8,13217
-dhisana/utils/generate_email_response.py,sha256=Xk6t2hW_QbumvXf5uUdsD-Lkq8hfzRD9QWioYcdWM1k,21194
+dhisana/utils/generate_email.py,sha256=cEIr9HbGduk-l_0dNY2lCBABFBmmcOHaMyI7oy8ChWQ,16470
+dhisana/utils/generate_email_response.py,sha256=LPsi2Orhb0W9SG8HDN2aA0w2pgW_c_HTDFAlw9R1cns,21391
 dhisana/utils/generate_flow.py,sha256=QMn6bWo0nH0fBvy2Ebub1XfH5udnVAqsPsbIqCtQPXU,4728
 dhisana/utils/generate_leads_salesnav.py,sha256=FG7q6GSm9IywZ9TgQnn5_N3QNfiI-Qk2gaO_3GS99nY,12236
-dhisana/utils/generate_linkedin_connect_message.py,sha256=QxsxDiT-3eQOqAAbW13d0HGJXV36WYPvC-7Zsw_2VTI,10208
+dhisana/utils/generate_linkedin_connect_message.py,sha256=niKnL1YwtROVsPctaIQFHeSIvr4ewqEougUBAvMmoC0,13385
 dhisana/utils/generate_linkedin_response_message.py,sha256=mWoSs5p2JSTIoFZFGm86x1kgs67J7dHPvGKZPzcdGdU,14569
 dhisana/utils/generate_structured_output_internal.py,sha256=gJ6Bdwz2VexL505crMOgYYHOhmZo0MCghRoDkLP62WM,33654
 dhisana/utils/google_custom_search.py,sha256=5rQ4uAF-hjFpd9ooJkd6CjRvSmhZHhqM0jfHItsbpzk,10071
@@ -57,7 +57,7 @@ dhisana/utils/lusha_tools.py,sha256=MdiWlxBBjSNpSKz8rhNOyLPtbeh-YWHgGiUq54vN_gM,
 dhisana/utils/mailgun_tools.py,sha256=brOgfEx-ciqEdDkXEfzMBfXkG0kRWVscg76tQDXb_lk,5826
 dhisana/utils/mailreach_tools.py,sha256=uJ_gIcg8qrj5-k3jnYYhpwLVnQncoA1swzr5Jfkc1JU,3864
 dhisana/utils/microsoft365_tools.py,sha256=ClqBzTrJ2SZM5K9nsOFyyHRfV-d-6jlxXNpNONtgLlY,18596
-dhisana/utils/openai_assistant_and_file_utils.py,sha256=-eyPcxFvtS-DDtYQGle1SU6C6CuxjulVIojFy27HeWc,8957
+dhisana/utils/openai_assistant_and_file_utils.py,sha256=snaawY7Y94q-XMzjK7UUylLy9fWlcmpWSOKcu1u04LE,9507
 dhisana/utils/openai_helpers.py,sha256=ZK9S5-jcLCpiiD6XBLkCqYcNz-AGYmO9xh4e2H-FDLo,40155
 dhisana/utils/openapi_spec_to_tools.py,sha256=oBLVq3WeDWvW9O02NCvY8bxQURQdKwHJHGcX8bC_b2I,1926
 dhisana/utils/parse_linkedin_messages_txt.py,sha256=g3N_ac70mAEuDDQ7Ott6mkOaBwI3ZvcsJD3R9RlYwPQ,3320
@@ -95,8 +95,8 @@ dhisana/workflow/agent.py,sha256=esv7_i_XuMkV2j1nz_UlsHov_m6X5WZZiZm_tG4OBHU,565
 dhisana/workflow/flow.py,sha256=xWE3qQbM7j2B3FH8XnY3zOL_QXX4LbTW4ArndnEYJE0,1638
 dhisana/workflow/task.py,sha256=HlWz9mtrwLYByoSnePOemBUBrMEcj7KbgNjEE1oF5wo,1830
 dhisana/workflow/test.py,sha256=E7lRnXK0PguTNzyasHytLzTJdkqIPxG5_4qk4hMEeKc,3399
-dhisana-0.0.1.dev272.dist-info/METADATA,sha256=gcVzqziXrrxkS44GPUwFkkf1igjYwMWE3PmWwxo_mT0,1190
-dhisana-0.0.1.dev272.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-dhisana-0.0.1.dev272.dist-info/entry_points.txt,sha256=jujxteZmNI9EkEaK-pOCoWuBujU8TCevdkfl9ZcKHek,49
-dhisana-0.0.1.dev272.dist-info/top_level.txt,sha256=NETTHt6YifG_P7XtRHbQiXZlgSFk9Qh9aR-ng1XTf4s,8
-dhisana-0.0.1.dev272.dist-info/RECORD,,
+dhisana-0.0.1.dev274.dist-info/METADATA,sha256=sPREfHEIvxxGk24-YL0zauwjU1n_pAUvaVmyLL_XKj0,1190
+dhisana-0.0.1.dev274.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+dhisana-0.0.1.dev274.dist-info/entry_points.txt,sha256=jujxteZmNI9EkEaK-pOCoWuBujU8TCevdkfl9ZcKHek,49
+dhisana-0.0.1.dev274.dist-info/top_level.txt,sha256=NETTHt6YifG_P7XtRHbQiXZlgSFk9Qh9aR-ng1XTf4s,8
+dhisana-0.0.1.dev274.dist-info/RECORD,,