OTHER RESEARCH SYSTEMS & COMPONENTS
On-device literature retrieval and evidence review
A retrieval-augmented generation (RAG) system that runs locally on a standard laptop to answer questions about experimental design, findings and citations. Evidence is organised by experiment, with source locations attached to answers and gaps stated explicitly. This project preceded the Biopharma Investment Research Library (BIR), which adopted its approach to organising evidence.
Results & validation
The September test used MiniCPM5-2B, a 2-billion-parameter model with 4-bit quantisation, on an RTX 3050 laptop GPU with 4 GB of memory. The retrieval index contained 507 papers and 38,089 passages.
In a retrieval test on a cancer-drug study, the system brought together five details from the main text and figure legends: established brain metastases, treatment from day 11 after implantation, selpercatinib, tumour burden measured by bioluminescence imaging (BLI), and no significant treatment benefit. It kept these separate from another experiment in the same paper that began pralsetinib on day 2.
Three further diagnostic questions exposed specific gaps: one located two endpoints but could not establish direction, one lacked a patient-level analysis, and one left a direct citation link unconfirmed. These tests used questions with known answers. The results describe performance on those cases, not overall accuracy.
Implementation
Document structure and source locations
Papers are parsed into sections, passages and figure legends. Semantic and keyword search can cover the full library or selected papers. Each piece of evidence retains a page or passage location for source review. Papers without full text are flagged and excluded from answer generation.
Evidence organised by experiment
A paper may report several experiments, with details spread across methods, results and figure legends. The system identifies which experiment each detail belongs to, then groups dosing time, intervention, measurements and results accordingly. This reduces the risk of mixing evidence from different experiments in one answer.
Evidence status
For each question, the system tracks the information needed and distinguishes confirmed source evidence, candidates and missing details. Answers use confirmed evidence and identify what has not been found.
Direct citation checks
Direct citations are identified from reference lists and in-text citation markers. Papers with only a topical connection are labelled separately.
Task allocation for a small model
A second review of relations by the same model was correct in only two of six known-answer cases and consumed part of the retrieval budget. It was replaced with programmatic checks of sources and direction. The model handles query planning and evidence selection, while code checks source references and experiment identity. The complete workflow runs with a 2-billion-parameter model on 4 GB of GPU memory.