Update utils.py

utils.py

@@ -498,11 +498,9 @@ def run_research_agent(
       1) Tavily search (up to 20 URLs).
       2) Firecrawl scrape => combined text
       3) Truncate to 12k tokens total
-      4) Split => at most 2 x 6k chunks => Summarize each chunk once => summaries
+      4) Split into chunks (each 4500 tokens) => Summarize each chunk individually => summaries
       5) Single final merge => final PDF
-      => 2 or 3 total LLM calls => drastically fewer calls => less chance of 429
-
-    Logs at each step for clarity.
+      => 2 or more total LLM calls (but no more than 10) to reduce the chance of rate limit errors.
     """
     print(f"[LOG] Starting LOW-CALL research agent for topic: {topic}")
 
@@ -547,17 +545,19 @@ def run_research_agent(
         print("[LOG] Step 3: Truncating combined text to 12,000 tokens if needed.")
         combined_content = truncate_text_tokens(combined_content, max_tokens=12000)
 
-        # Step 4: At most 2 chunks => Summaries
-        print("[LOG] Step 4: Splitting text into up to 2 chunks (6,000 tokens each). Summarizing each chunk.")
+        # Step 4: Splitting text into chunks (4500 tokens each) and summarizing each chunk.
+        print("[LOG] Step 4: Splitting text into chunks (4500 tokens each). Summarizing each chunk.")
         tokenizer = tiktoken.get_encoding("cl100k_base")
         tokens = tokenizer.encode(combined_content)
-        chunk_size = 6000
-
-        groq_client = Groq(api_key=os.environ.get("GROQ_API_KEY"))
+        chunk_size = 4500          # Reduced chunk size to avoid exceeding the LLM's TPM limit.
+        max_chunks = 10            # Allow up to 10 chunks (and thus 10 LLM calls).
         summaries = []
         start = 0
         chunk_index = 1
-        while start < len(tokens):
+
+        groq_client = Groq(api_key=os.environ.get("GROQ_API_KEY"))
+
+        while start < len(tokens) and chunk_index <= max_chunks:
             end = min(start + chunk_size, len(tokens))
             chunk_text = tokenizer.decode(tokens[start:end])
             print(f"[LOG] Summarizing chunk {chunk_index} with ~{len(tokens[start:end])} tokens.")
@@ -566,7 +566,6 @@ You are a specialized summarization engine. Summarize the following text
 for a professional research report. Provide accurate details but do not 
 include chain-of-thought or internal reasoning. Keep it concise, but 
 include key data points and context:
-
 {chunk_text}
 """
             data = {
@@ -578,12 +577,8 @@ include key data points and context:
             response = call_llm_with_retry(groq_client, **data)
             summary_text = response.choices[0].message.content.strip()
             summaries.append(summary_text)
-
             start = end
             chunk_index += 1
-            # Because chunk_size=6000, only 2 chunks max
-            if chunk_index > 2:
-                break
 
         # Step 5: Single final merge call
         print("[LOG] Step 5: Doing one final merge of chunk summaries.")
@@ -604,20 +599,16 @@ Produce a long, academic-style research paper with the following structure:
 - Footnotes or inline citations referencing the URLs
 - Conclusion
 - References / Bibliography (list these URLs at the end)
-
 Requirements:
 - Minimal bullet points, prefer multi-paragraph
 - Each section at least 2-3 paragraphs
 - Aim for 1500+ words if possible
 - Under 6000 tokens total
 - Professional, academic tone
-
 Partial Summaries:
 {merged_input}
-
 References (URLs):
 {references_text}
-
 Now, merge these partial summaries into one thoroughly expanded research paper:
 """
         final_data = {