Feature Projects
Python RAG Service — Documentation Retrieval & AI Infrastructure
Implemented and tested | View project on GitHub | Explore the documentation
I built a platform-agnostic Python RAG service that turns documentation into searchable evidence for grounded AI answers. The service ingests and normalizes content, preserves document structure through heading-aware chunking, combines semantic and keyword retrieval, and generates answers with citations to retrieved sources.
The first implementation uses WordPress as client and connector. The architecture separates the content connector, indexing pipeline, retrieval system, answer-generation layer, API, and client so the core service can support other sources and applications.
Retrieval quality was a central engineering concern. I combined vector retrieval and BM25 keyword retrieval using Reciprocal Rank Fusion, added Voyage AI reranking, and implemented a configurable support gate that can return no results when retrieved evidence falls below the cutoff. Separate Search and Answer APIs let clients retrieve documentation with or without generating an answer.
Documentation engineering work included:
- Designing a platform-neutral content model, retrieval metadata, and stable document and chunk identifiers
- Implementing HTML normalization, heading-aware chunking, and incremental indexing
- Building hybrid retrieval with Voyage embeddings, Qdrant, BM25, ranking fusion, and reranking
- Implementing grounded answer generation and validating citation identifiers against supplied sources
- Exposing authenticated FastAPI endpoints and an OpenAPI-based API reference
- Building a WordPress reference client with Search and Ask interfaces
- Creating versioned evaluation cases, saved reports, and failure analysis for retrieval, answer structure, citation behavior, and unsupported queries
- Building the service’s documentation system with content models, templates, terminology rules, authoring tools, automated validation, and CI publishing
The October 22, 2026 evaluation snapshot passed 21 of 22 retrieval cases, returned no chunks for all eight expected-empty queries, and passed answer-structure checks for all 22 cases. The documented retrieval failure involved incomplete source coverage for a question requiring multiple sections. These results describe a defined benchmark on one corpus; structural checks and qualitative answer review assess different aspects of quality.
The project also includes a separate operations runner and n8n workflows for scheduled indexing, retrieval evaluation, quality checks, Jira regression reporting, and failure alerts.
The completed project connects documentation structure, retrieval behavior, developer experience, and ongoing operations in one working system.
View the developer documentation case study →
DevEx Cohere API Documentation
Cohere publishes its developer documentation from a public, open-source repository. I forked that repository and used it as the foundation for a documentation engineering project. The project included auditing the existing developer experience, redesigning the information architecture, improving Guides and Concepts and API reference content, filling identified documentation gaps, and building standards and automation to make those improvements repeatable.
The changes publish to a live documentation site, so the work can be evaluated as a functioning documentation experience rather than as isolated writing samples.
The project also became an early testing ground for AI-assisted documentation workflows. I created documentation standards that AI tools can work from—including a content model, writing style guide, platform rules, and page-type templates—and reusable AI skills for authoring, reviewing, and updating documentation.
The goal of this work is not simply to use AI to generate content. It is to define how AI participates in the documentation workflow, what standards govern its work, what context it needs, and where human judgment and approval remain part of the process.
The standards and workflow patterns explored here informed the content model, templates, authoring guidance, and automated validation implemented for the Python RAG Service documentation.