Chunking architecture
H2- and H3-anchored sections at roughly 300 tokens, each one self-contained enough to stand alone as an answer. Enriched with product, version, and topic-type metadata so retrieval can filter before it ranks.
chunk 01type conceptid thesis
Two audiences, one source
Documentation now has two readers: the developer on the page, and the model retrieving from it. They fail differently. A person skims past a bad heading; a retriever returns the wrong chunk and the model answers confidently from it.
So I architect every layer for both: structure, metadata, chunking, and evaluation. I treat retrieval quality as a number I can move, not a feeling. Fourteen years to get here: DITA and single-source publishing, information architecture at Adobe HelpX scale, and now retrieval engineering at Actian.
The search box above is the argument. It indexes this page at load, scores your query with term frequency weighted by field, and ranks the sections. Same shape as the pipeline, minus the embeddings.
chunk 02type referenceid retrieval-stack
What I actually build
Retrieval
Evaluation
RAGAS baselines
Answer relevance, faithfulness, and context recall measured against a repeatable baseline. Evaluation is the part most RAG projects skip, and the reason most of them cannot be trusted in production.
Model selection
Embedding benchmarks
sentence-transformers, text-embedding-3-small, and bge-m3 scored against the actual documentation corpus rather than a public leaderboard. The corpus decides, not the ranking table.
Interfaces
MCP servers
Exposing a corpus, its analytics, and its build status as tools any MCP client can discover and call. Static pages become a live source an assistant can query mid-conversation.
Graphs
Typed knowledge graphs
Documentation modelled as nodes and typed edges (REQUIRES, COVERS, BELONGS_TO) so retrieval can follow relationships instead of returning isolated chunks.
Feedback
Content-gap analytics
A GA4 to BigQuery pipeline surfacing zero-result searches and high-exit pages weekly. The corpus gets more answerable over time rather than just larger.
chunk 03type indexid work
Shipped, and readable
MCP server · TypeScript
docs-mcp
Five tool handlers exposing a documentation corpus, GA4 content gaps, and Jenkins CI status to Claude in one conversation. Runs with no credentials: mock fixtures, documented upgrade paths to pgvector and BigQuery.
View repository →MCP server · evaluation
knowflow
The evolution of docs-mcp: adds a RAGAS-style evaluation loop measuring relevance, faithfulness, and recall, so retrieval quality is measured rather than assumed. Retrieval plus the loop that tells you it works.
View repository →RAG pipeline · Python
Documentation-AI-Assistant
A retrieval-augmented assistant grounding LLM answers in a documentation corpus through embeddings and semantic search, with retrieval guardrails on what it is allowed to answer from.
View repository →Knowledge graph · Python
knowledge-graphs-for-ia
Turns a documentation set into a typed knowledge graph and retrieves across relationships rather than isolated chunks. Information architecture made executable. GraphRAG generation is in progress.
View repository →Pipeline · CI gate
docs-style-guard
Writing standards as an automated check that runs on every change, callable over MCP. A style guide nobody reads becomes a gate nothing gets past.
View repository →Practice · writing
Knowlayer
My independent practice and published work on enterprise content retrievability: chunk integrity, knowledge graphs, and the shift from pageview KPIs to AI-readiness metrics.
Visit Knowlayer →chunk 04type timelineid experience
Fourteen years
2024 to nowBengaluru, India
Principal Information Architect
Actian
Leading documentation engineering across the enterprise data products, and building the retrieval layer that makes the corpus queryable by AI systems.
- Built a production RAG chatbot grounded in the corpus, using embeddings, vector search, and retrieval guardrails.
- Introduced RAGAS evaluation covering answer relevance, faithfulness, and context recall, establishing a repeatable baseline for retrieval changes.
- Built an MCP server exposing the docs corpus, GA4 analytics, and CI status to Claude through natural-language tool calls.
- Designed H2/H3-anchored 300-token chunking with product, version, and topic-type metadata for filtered retrieval.
- Built a GA4 to BigQuery pipeline surfacing zero-result searches as a weekly content-gap dashboard.
- Implemented Jenkins CI/CD with automated validation gates, cutting build and release time by 40 percent.
- Rebuilt the documentation portal on docs-as-code, and led an SEO programme delivering 50 percent organic traffic growth in six months.
2021 to 2024Bengaluru, India
Information Architect
Adobe
Content strategy and information architecture for Adobe HelpX, serving over a million users per quarter across developer and enterprise documentation.
- Built M&A documentation frameworks adopted across multiple acquisitions, cutting engineering onboarding time by 70 percent.
- Implemented OpenAPI documentation with an interactive testing environment, among the top three developer requests.
- Led AI-assisted content migration, taxonomy design, and structured authoring to improve reuse across acquired ecosystems.
2018 to 2021Bengaluru, India
Senior Technical Writer / Information Architect
ABB
Documentation for a full developer ecosystem: API guides, CLI references, and Swift and Kotlin mobile SDKs.
- Restructured the documentation library using card sorting and tree testing, improving navigation satisfaction by 40 percent.
- Led the DITA migration from legacy Word and InDesign sources, enabling multi-channel publishing.
2012 to 2018Gurgaon, India
Senior Technical Writer
Aristocrat Technologies
Structured content foundations: DITA information modelling, XML publishing, and enterprise taxonomy.
- Designed DITA information models and custom XSLT and XSL-FO plug-ins for PDF, HTML, and CHM from a single source.
- Developed enterprise taxonomy, controlled vocabularies, and XML validation rules.
chunk 05type referenceid skills
Tools
AI & retrieval
Engineering
Information architecture
Data & APIs
chunk 06type dataid results
Numbers I can source
+45%
Discoverability
From a standardised IA and navigation model at Actian
+50%
Organic traffic
Metadata, URL structure, schema markup, and internal linking, in six months
−70%
Onboarding time
Through the M&A documentation framework at Adobe
−40%
Build & release time
Jenkins pipeline with automated validation gates
+40%
API self-service
REST and GraphQL reference with interactive testing
1M+
Quarterly users
Adobe HelpX, the corpus the IA served