Microservice for using chained LLM calls and document embeddings to explore sets of documents. Generates a prompt and source metadata to be passed to an LLM pipeline for question answering.
reduce latency-> build_docstore() runtime now down to < 0.5s (~.3s) w/ Redis cachingcache embeddings of split docsuse as key Arxiv ID and store array of b64 str representations of embeddings
test with mulitple documentsfix occasional bug with returning "sources" w/ answer (source sometimes changes for same answer/query)
{
'prompt': 'Using the context below...',
'sources':[
{'source': {'section': 'Introduction', 'content': 'The long-sought Higgs boson(s) $h$ of electroweak symmetry breaking\\nin particle physics may soon be observed at the CERN Large Hadron\\nCollider (LHC) through the diphoton decay mode ($h\\\\rightarrow\\\\gamma\\\\gamma$).\\nPurely hadronic standard model processes are a copious source of diphotons,\\nand a narrow Higgs boson signal at relatively low masses will appear\\nas a small peak above this considerable background. A precise theoretical\\nunderstanding of the kinematic distributions for diphoton production\\nin the standard model could provide valuable guidance in the search\\nfor the Higgs boson signal and assist in the important measurement \\nof Higgs boson coupling strengths.', 'title': 'Calculation of prompt diphoton production cross sections at Tevatron and LHC energies', 'id': '0704.0001'}}...
]
}