Source adapters
The project documents a generic source-adapter contract, with Dropbox API extraction treated as the first private adapter rather than the public boundary.
Project Brief
Content Intelligence
An artifact-to-RAG pipeline that standardizes text, transcript-derived, and OCR-derived content into provenance-rich records and serves cited answers through hybrid retrieval.
Live Demo
Open the standalone retrieval workspace
Open the standalone surface to see the published corpus status, ask a question, and inspect citations without scrolling through the project narrative.
Implemented Pipeline
From source adapters to RAG-ready evidence
Useful decisions often live in unstructured notes, PDFs, transcripts, and planning records. The pipeline turns approved inputs into normalized records while retaining source identity, checksums, citation labels, and review status through every downstream step.
The public implementation uses synthetic text, transcript-derived, and OCR-derived examples to demonstrate conversion, segmentation, safety review, retrieval, and cited answers without publishing private source material.
Inspect first
- Live cited retrieval over the reviewed release corpus
- Source adapter and information-object contracts
- Corpus fingerprint and public-safety review output
- Deterministic retrieval evaluations and sample report
Evidence
What the workflow proves
Conversion and normalization
Transcript and OCR-style examples become cleaned, citation-preserving corpus records after public-safety review.
Retrieval-ready index
Generated RAG index records add stable chunk IDs, source metadata, method/evidence collections, citation labels, safety levels, embedding metadata, and retrieval fields.
Evaluated boundary
The validation step checks generated objects and requires a passing public-safety review before records can be treated as retrieval-ready.
Serving Layer
Hybrid retrieval with a safe fallback
The workflow produces Vectorize-ready records. With Cloudflare bindings configured, the portfolio Worker combines vector and lexical retrieval; otherwise, the demo uses lexical retrieval over the same safety-reviewed records.
Inspect in source
Safety Boundary
Source-grounded, source-limited
Public examples must use synthetic, sanitized, public-domain, or clearly licensed source material. Raw private artifacts, private source paths, credentials, course identifiers, and copyrighted lecture text stay outside the public repository and public RAG index.