bakedin.co

Mark license-verified papers as included in the training corpus.

Extract PICO-structured claims from papers using Bedrock Haiku, then compute GRADE-lite evidence grades deterministically.

Build a CPT training corpus JSONL from extracted papers.

Run Bedrock Haiku classification on unclassified figures.

Papers that are license-verified and ready for corpus inclusion.

CHUNK-level semantic search — returns the actual paragraphs of body text most relevant to your query, with their paper context.

Gap analysis and finish-line cost projection.

BEST general-purpose search: fuses corpus_keyword_search + corpus_semantic_search via Reciprocal Rank Fusion (RRF).

Ingest course_content_graph edges from S3-hosted JSONL.

Ingest staged papers + chunks from S3-hosted JSONL into corpus_papers + corpus_section_chunks.

Full-text keyword search over the corpus (172K papers).

Meaning-based search over the corpus via pgvector cosine similarity.

Current corpus state across all aspects: paper intake, pipeline stages, derived artifacts (claims/SFT/figures/safety/cards), full-text sections, books, regulatory/tribal/underwriting content, quality and relevance distributions, and empirical cost history.

Compares the latest corpus snapshot to one from N days ago, showing deltas on headline metrics (papers, abstracts, license-verified, CPT tokens, claims, SFT pairs + source papers, paper sections, books).

Assemble the cross-axis 'Anatomy of a Datacenter': a facility's scale + siting (infra_projects), the energy build-out in its state, the subsidies/cost-shift in its jurisdiction (infra_subsidies), and the corpus papers explaining the underlying constraints.

Find and report duplicate papers across different sources (matching DOI or ArXiv ID).

Fast metadata-only indexing.

Stats on paper coverage per domain and overall progress toward the 300+ paper corpus target.

Download PDFs and extract figures for papers.

Download PDF, extract full text with section parsing (IMRaD), chunk for training.

Extract text from papers that have abstracts or PDF URLs.

Search and browse extracted figures with filters.

Get full metadata for a specific figure by figure_id.

Generate SFT training Q&A pairs from extracted claims using Bedrock Haiku.

Run full harvest across all domains and all sources.

Harvest papers for a specific domain from a source API.

Queue-state invariants for the ingest pipeline.

Full details for a single paper by paper_id (e.g.

Query the paper registry with optional filters.

Pipeline processing progress across all domains and stages.

Verify licenses and promote high-quality discovered papers.

Mark a paper as rejected with a reason.

View the audit log of papers rejected during harvesting.

Run canary test: preflight checks, process 5 papers through full pipeline, run all validators.

Batch-verify license status for papers in 'discovered' status.