Update app.py

app.py

@@ -35,58 +35,58 @@ print(f"✅ Loaded on {device.upper()}")
 # SYSTEM PROMPT
 # =========================
 
-SYSTEM_PROMPT = """You are a memory compression engine. Compress and merge facts into one short dense paragraph.
+SYSTEM_PROMPT = """You are a memory compression engine. Compress and merge facts into dense paragraphs.
 
 EXAMPLE 1:
 EXISTING MEMORY: (none)
 USER SAID: I am building a weather app using React and OpenWeatherMap API.
 ASSISTANT REPLIED: Fetch data with axios. Store API key in .env via process.env.
-UPDATED MEMORY: User building React weather app using OpenWeatherMap API. Data fetched via axios. API key stored in .env.
+UPDATED MEMORY: User building React weather app using OpenWeatherMap API. Data fetched via axios. API key stored in .env via process.env.
 
 EXAMPLE 2:
-EXISTING MEMORY: User building React weather app using OpenWeatherMap API. Data fetched via axios. API key stored in .env.
+EXISTING MEMORY: User building React weather app using OpenWeatherMap API. Data fetched via axios. API key stored in .env via process.env.
 USER SAID: How do I cache the weather data so I do not hit the API limit?
 ASSISTANT REPLIED: Use localStorage to cache responses with a timestamp. If cache is under 10 minutes old, return it instead of calling the API.
-UPDATED MEMORY: User building React weather app using OpenWeatherMap API. Data fetched via axios, cached in localStorage with 10-minute expiry to avoid API limit.
+UPDATED MEMORY: User building React weather app using OpenWeatherMap API. Data fetched via axios. API key in .env. Responses cached in localStorage with 10-minute timestamp expiry to avoid API rate limit.
 
 EXAMPLE 3:
 EXISTING MEMORY: User building job board with Django, React, PostgreSQL. JWT auth via djangorestframework-simplejwt. Custom user model with company and jobseeker roles. Job model has title, description, skills, salary range, location.
 USER SAID: How do job seekers apply for a job?
-ASSISTANT REPLIED: Create Application model with ForeignKey to Job and User, status field, resume FileField in S3.
-UPDATED MEMORY: User building job board with Django, React, PostgreSQL, JWT auth. Custom user model with company/jobseeker roles. Job model has title, description, skills, salary range, location. Application model has ForeignKey to Job and User, status field, resume stored in S3.
+ASSISTANT REPLIED: Create Application model with ForeignKey to Job and User, status field (applied, reviewed, rejected, accepted), resume FileField stored in S3.
+UPDATED MEMORY: User building job board with Django, React, PostgreSQL. JWT auth via djangorestframework-simplejwt. Custom user model with company and jobseeker roles. Job model has title, description, skills, salary range, location. Application model has ForeignKey to Job and User, status field (applied/reviewed/rejected/accepted), resume FileField stored in S3.
 
 EXAMPLE 4:
-EXISTING MEMORY: User building job board with Django, React, PostgreSQL, JWT auth. Custom user model with company/jobseeker roles. Job model has title, description, skills, salary range, location. Application model has ForeignKey to Job and User, status field, resume stored in S3.
+EXISTING MEMORY: User building job board with Django, React, PostgreSQL. JWT auth via djangorestframework-simplejwt. Custom user model with company and jobseeker roles. Job model has title, description, skills, salary range, location. Application model has ForeignKey to Job and User, status field (applied/reviewed/rejected/accepted), resume FileField stored in S3.
 USER SAID: I want to add search and filters for title, location, and salary range.
-ASSISTANT REPLIED: Use Django Q objects and django-filter. Add query params to job list endpoint.
-UPDATED MEMORY: User building job board with Django, React, PostgreSQL, JWT auth. Company/jobseeker roles. Job and Application models complete with S3 resumes. Job search via django-filter and Q objects on title, location, salary range.
+ASSISTANT REPLIED: Use Django Q objects and django-filter. Add query params title, location, salary_min, salary_max to job list endpoint.
+UPDATED MEMORY: User building job board with Django, React, PostgreSQL. JWT auth via djangorestframework-simplejwt. Company and jobseeker roles. Job model has title, description, skills, salary range, location. Application model has status field and S3 resume. Job search and filtering via django-filter and Q objects on title, location, salary_min, salary_max query params.
 
 EXAMPLE 5:
-EXISTING MEMORY: User building job board with Django, React, PostgreSQL, JWT auth. Company/jobseeker roles. Job and Application models complete with S3 resumes. Job search via django-filter and Q objects on title, location, salary range.
-USER SAID: How do I notify applicants when status changes?
-ASSISTANT REPLIED: Use Django signals on Application post_save. Trigger SendGrid email via Celery async task.
-UPDATED MEMORY: User building job board with Django, React, PostgreSQL, JWT auth, Celery, SendGrid. Company/jobseeker roles. Job and Application models with S3 resumes and django-filter search. Status change notifications via Django signals and Celery tasks.
+EXISTING MEMORY: User building job board with Django, React, PostgreSQL. JWT auth via djangorestframework-simplejwt. Company and jobseeker roles. Job model has title, description, skills, salary range, location. Application model has status field and S3 resume. Job search and filtering via django-filter and Q objects on title, location, salary_min, salary_max query params.
+USER SAID: How do I notify applicants when their application status changes?
+ASSISTANT REPLIED: Use Django signals. On Application post_save, detect status change and trigger email via Celery async task using SendGrid.
+UPDATED MEMORY: User building job board with Django, React, PostgreSQL. JWT auth via djangorestframework-simplejwt. Company and jobseeker roles. Job model has title, description, skills, salary range, location. Application model has status field (applied/reviewed/rejected/accepted) and S3 resume. Job search via django-filter and Q objects. Status change notifications triggered via Django signals on Application post_save, sending emails via Celery tasks and SendGrid.
 
 EXAMPLE 6:
-EXISTING MEMORY: User building job board with Django, React, PostgreSQL, JWT auth, Celery, SendGrid. Company/jobseeker roles. Job and Application models with S3 resumes and django-filter search. Status change notifications via Django signals and Celery tasks.
+EXISTING MEMORY: User building job board with Django, React, PostgreSQL. JWT auth via djangorestframework-simplejwt. Company and jobseeker roles. Job model has title, description, skills, salary range, location. Application model has status field (applied/reviewed/rejected/accepted) and S3 resume. Job search via django-filter and Q objects. Status change notifications triggered via Django signals on Application post_save, sending emails via Celery tasks and SendGrid.
 USER SAID: How do I deploy this on a VPS?
-ASSISTANT REPLIED: Docker Compose with Django, React, PostgreSQL, Redis, Celery services. Gunicorn behind nginx. Certbot for SSL.
-UPDATED MEMORY: User building job board with Django, React, PostgreSQL, JWT auth, Celery, SendGrid, Redis. Company/jobseeker roles. Job and Application models with S3 resumes and django-filter search. Status notifications via Django signals. Deployed via Docker Compose with Gunicorn, nginx, Certbot SSL.
+ASSISTANT REPLIED: Docker Compose with services for Django, React, PostgreSQL, Redis, Celery. Serve Django via Gunicorn behind nginx. Certbot for SSL. Secrets in .env file.
+UPDATED MEMORY: User building job board with Django, React, PostgreSQL, Redis, Celery, SendGrid. JWT auth via djangorestframework-simplejwt. Company and jobseeker roles. Job model has title, description, skills, salary range, location. Application model has status field and S3 resume. Job search via django-filter and Q objects. Status change emails via Django signals and Celery tasks. Deployed via Docker Compose with Gunicorn, nginx reverse proxy, Certbot SSL, secrets in .env.
 
 EXAMPLE 7:
-EXISTING MEMORY: User building job board with Django, React, PostgreSQL, JWT auth, Celery, SendGrid, Redis. Company/jobseeker roles. Job and Application models with S3 resumes and django-filter search. Status notifications via Django signals. Deployed via Docker Compose with Gunicorn, nginx, Certbot SSL.
+EXISTING MEMORY: User building job board with Django, React, PostgreSQL, Redis, Celery, SendGrid. JWT auth via djangorestframework-simplejwt. Company and jobseeker roles. Job model has title, description, skills, salary range, location. Application model has status field and S3 resume. Job search via django-filter and Q objects. Status change emails via Django signals and Celery tasks. Deployed via Docker Compose with Gunicorn, nginx reverse proxy, Certbot SSL, secrets in .env.
 USER SAID: What is still left to build?
-ASSISTANT REPLIED: Admin panel, pagination, rate limiting, frontend loading states and error handling.
-UPDATED MEMORY: User building job board with Django, React, PostgreSQL, JWT auth, Celery, SendGrid, Redis. Company/jobseeker roles. Job, Application models with S3 resumes, django-filter search, Docker Compose deployment. Pending: admin panel, pagination, rate limiting, frontend loading states and error handling.
+ASSISTANT REPLIED: Admin panel for moderating posts, pagination on job listings, rate limiting on API, frontend loading states and error handling.
+UPDATED MEMORY: User building job board with Django, React, PostgreSQL, Redis, Celery, SendGrid. JWT auth via djangorestframework-simplejwt. Company and jobseeker roles. Job model has title, description, skills, salary range, location. Application model has status field and S3 resume. Job search via django-filter and Q objects. Status change emails via Django signals and Celery tasks. Deployed via Docker Compose with Gunicorn, nginx, Certbot SSL. Pending: admin moderation panel, pagination, API rate limiting, frontend loading states and error handling.
 
 STRICT RULES:
 - Output ONLY the updated memory. No labels. No preamble. No explanation.
-- COMPRESS the existing memory. Do not copy it verbatim. Rewrite it shorter.
-- Keep ALL technical facts. Remove only filler words.
-- Add new facts merged in, not appended as separate sentences.
-- No filler: no "ensuring", "enhances", "this setup", "this approach", "in order to".
+- COMPRESS the existing memory. Do not copy it verbatim. Rewrite it shorter and denser.
+- Keep ALL technical facts: stack, frameworks, APIs, models, field names, architecture decisions, unfinished tasks, user preferences.
+- Add new facts merged in naturally, not appended as separate sentences.
+- No filler: no "ensuring", "enhances", "this setup", "this approach", "in order to", "it is worth noting".
 - No questions. No advice. No "you". No "I".
-- One short dense paragraph. Maximum 3 sentences."""
+- Dense technical paragraph. Maximum 8 sentences."""
 
 # =========================
 # FILLER PATTERNS
@@ -101,63 +101,84 @@ FILLER_PATTERNS = [
     r"for (better|improved|efficient|effective|optimal)\s[^.]*\.",
     r"in order to\s[^.]*\.",
     r"To (enhance|improve|ensure|enable)\s[^.]*\.",
+    r"It is worth noting that\s[^.]*\.",
+    r"Additionally,\s*(it|this)\s[^.]*\.",
 ]
 
+# =========================
+# MEMORY LIMIT CONFIG
+# =========================
+
+MEMORY_SOFT_LIMIT = 1600   # ~400 tokens — compress aggressively beyond this
+MEMORY_HARD_LIMIT = 2000   # ~500 tokens — absolute cap, never exceed
+
+
 # =========================
 # HELPERS
 # =========================
 
 def clean_assistant_message(text: str) -> str:
     """
-    Strip code blocks from assistant responses.
-    Extract function/class names and key terms before removing.
-    Keep only prose explanation, cap at 500 chars.
+    Strip code blocks and inline code backticks from assistant responses.
+    Model picks up function/class names from prose naturally.
+    Cap at 800 chars to give model more context from long responses.
     """
-    # Extract key identifiers from code before removing
-    code_blocks = re.findall(r"```[\w]*\n?(.*?)```", text, re.DOTALL)
-    extracted_terms = []
-
-    for block in code_blocks:
-        # Grab function/class/variable names
-        names = re.findall(
-            r"(?:def|class|const|let|var|function)\s+(\w+)", block
-        )
-        extracted_terms.extend(names)
-
-    # Remove code blocks
+    # Remove full code blocks entirely
     text = re.sub(r"```[\w]*\n?.*?```", "", text, flags=re.DOTALL)
 
-    # Remove inline code but keep the text
+    # Remove inline code backticks but keep the text inside
     text = re.sub(r"`([^`]+)`", r"\1", text)
 
-    # Append extracted key names if any
-    if extracted_terms:
-        text += " Key identifiers: " + ", ".join(extracted_terms) + "."
-
     # Collapse whitespace
     text = re.sub(r"\s{2,}", " ", text).strip()
 
-    return text[:500]
+    return text[:800]
 
 
-def enforce_memory_limit(text: str, max_chars: int = 600) -> str:
+def enforce_memory_limit(text: str) -> str:
     """
-    Hard cap on memory length.
-    If over limit, keep complete sentences up to the limit.
+    Three-stage memory length enforcement.
+
+    Stage 1 — Under 1600 chars (~400 tokens):
+        Memory is healthy. Return as-is.
+
+    Stage 2 — Between 1600 and 2000 chars (soft limit):
+        Memory is getting long. Keep complete sentences
+        that fit within 2000 chars. Oldest appended facts
+        may be trimmed; core stack in early sentences is preserved.
+
+    Stage 3 — Over 2000 chars (hard limit):
+        Force trim to last complete sentence before 2000 chars.
+        Never cuts mid-sentence.
     """
-    if len(text) <= max_chars:
+    # Stage 1 — healthy
+    if len(text) <= MEMORY_SOFT_LIMIT:
         return text
 
-    sentences = re.split(r"(?<=[.!?])\s+", text)
-    result = ""
+    # Stage 2 — soft limit: trim to complete sentences within hard limit
+    if len(text) <= MEMORY_HARD_LIMIT:
+        sentences = re.split(r"(?<=[.!?])\s+", text)
+        result = ""
+        for sentence in sentences:
+            candidate = (result + " " + sentence).strip()
+            if len(candidate) <= MEMORY_HARD_LIMIT:
+                result = candidate
+            else:
+                break
+        return result.strip()
+
+    # Stage 3 — hard limit: force trim at last period before 2000 chars
+    trimmed = text[:MEMORY_HARD_LIMIT]
+    last_period = trimmed.rfind(".")
+    if last_period != -1:
+        trimmed = trimmed[:last_period + 1]
+
+    return trimmed.strip()
 
-    for sentence in sentences:
-        if len(result) + len(sentence) + 1 <= max_chars:
-            result += ("" if not result else " ") + sentence
-        else:
-            break
 
-    return result.strip()
+def strip_backticks(text: str) -> str:
+    """Remove any backtick formatting that leaks into memory output."""
+    return re.sub(r"`([^`]+)`", r"\1", text)
 
 # =========================
 # REQUEST MODEL
@@ -189,6 +210,10 @@ UPDATED MEMORY:"""
         {"role": "user",   "content": user_content},
     ]
 
+    # =========================
+    # FORMAT CHAT
+    # =========================
+
     text = tokenizer.apply_chat_template(
         messages,
         tokenize=False,
@@ -200,14 +225,22 @@ UPDATED MEMORY:"""
         return_tensors="pt"
     ).to(model.device)
 
+    # =========================
+    # GENERATE
+    # =========================
+
     output = model.generate(
         **inputs,
-        max_new_tokens=220,
+        max_new_tokens=400,
         do_sample=False,
         repetition_penalty=1.15,
         eos_token_id=tokenizer.eos_token_id,
     )
 
+    # =========================
+    # DECODE
+    # =========================
+
     result = tokenizer.decode(
         output[0][inputs.input_ids.shape[1]:],
         skip_special_tokens=True
@@ -236,6 +269,12 @@ UPDATED MEMORY:"""
     for pattern in FILLER_PATTERNS:
         result = re.sub(pattern, "", result, flags=re.IGNORECASE)
 
+    # =========================
+    # CLEAN — strip backticks
+    # =========================
+
+    result = strip_backticks(result)
+
     # =========================
     # CLEAN — deduplicate lines
     # =========================
@@ -255,7 +294,7 @@ UPDATED MEMORY:"""
     # HARD MEMORY LENGTH CAP
     # =========================
 
-    result = enforce_memory_limit(result, max_chars=600)
+    result = enforce_memory_limit(result)
 
     return {"memory": result}
 
@@ -268,8 +307,14 @@ def root():
     return {
         "status": "Memory Summarizer Running 🚀",
         "model": MODEL_ID,
-        "device": device.upper()
+        "device": device.upper(),
+        "memory_soft_limit": f"{MEMORY_SOFT_LIMIT} chars (~400 tokens)",
+        "memory_hard_limit": f"{MEMORY_HARD_LIMIT} chars (~500 tokens)",
     }
 
+# =========================
+# RUN
+# =========================
+
 if __name__ == "__main__":
     uvicorn.run("app:app", host="0.0.0.0", port=7860)