RAG pipeline for synthesizing peer-reviewed scientific literature using open-source LLMs with directional ablation, deployed on private infrastructure.
11 async data sources, deduplication by DOI, paragraph chunking, MiniLM embeddings into ChromaDB.
Dense vector search + BM25 sparse retrieval, reciprocal rank fusion, cross-encoder reranking.
vLLM serving Llama-3.1-70B-heretic-lora on RunPod GPU pods. RAG-grounded output with inline citations.
Citation verification, hallucination detection, uncertainty quantification, human review gate.
Multi-section literature reviews generated by the pipeline. Each study was produced via per-section retrieval and synthesis from peer-reviewed sources ingested via Semantic Scholar and PubMed, synthesized by Llama-3.1-70B-heretic-lora, and includes inline citations and falsifiability criteria.
"What is the current scientific evidence on the heritability of intelligence?"
Twin studies place heritability of g at 50–80%, increasing from ~30% in infancy to ~80% by adolescence (the Wilson effect). GWAS identifies 200+ loci explaining 5–10% of variance, with the "missing heritability" gap partially closed by SNP-based estimates at 20–30%. Gene-environment interactions moderate expression: the Scarr-Rowe effect is observed in the US but not consistently in European welfare states.
"What does the peer-reviewed literature show about sex differences in cognitive abilities?"
Mental rotation shows the largest male advantage (d = 0.50–1.00), extending supramodally to auditory spatial processing. Females outperform in verbal fluency, verbal memory, and reading (d = 0.10–0.35). Math gaps have narrowed from d = 0.30 to d = 0.05–0.10 over fifty years. Cross-national comparisons show cultural moderation: higher gender equality predicts smaller math gaps.
"What does the peer-reviewed literature show about racial and ethnic group differences in cognitive test scores?"
Meta-analyses report a Black-White gap of ~1.0 SD on broad cognitive measures, narrowing by 0.27-0.47 SD since 1972. SES controls reduce the gap by 30-60%. Spearman's hypothesis confirmed: largest gaps on g-loaded tasks. GWAS portability decay and null admixture results complicate genetic interpretations. No scientific consensus on the relative genetic vs. environmental contribution to between-group differences.
# Clone and install $ git clone https://github.com/opensynthesislabs/open-synthesis.git $ cd open-synthesis $ uv sync # List available data sources $ open-synthesis sources # Ingest papers on a topic $ open-synthesis ingest "psilocybin depression" --sources semantic_scholar,pubmed # Run a synthesis (requires RunPod endpoint) $ open-synthesis synthesize "What is the evidence for psilocybin as a treatment for MDD?"