UPDATE_LOG.md

Update Log (Post-Hackathon)

12-03-2026 : Optimizations applied to app.py

To significantly improve the overall latency of the paper synthesis pipeline, the following performance optimizations were made in app.py today:

1. Model Warm-up on Startup:

   • Added an @app.on_event("startup") FastAPI hook.
   • The SentenceTransformer representation model is now pre-loaded into memory immediately when the server starts. This eliminates the heavy overhead (~80+ seconds) previously experienced on the very first search query.

2. Parallel Summarization & Claim Extraction:

   • Replaced the sequential per-paper summary and claim extraction logic with concurrent.futures.ThreadPoolExecutor.
   • Text processing now runs concurrently across background threads, effectively speeding up the enrichment stage of the pipeline.

3. In-Memory Endpoint Caching:

   • Introduced a global SEARCH_CACHE to store full API responses.
   • Search requests with identical queries (topic, max_results, similarity_threshold) will now return immediately via a direct cache hit, delivering a near-zero latency response without re-triggering network fetches or re-calculating graph elements.

4. Added Benchmarking Script:

   • Created benchmark.py for granular profiling of the search pipeline.