Documentation discovery

Semantic Documentation Search for Agentic Platforms

When an agent hits an unfamiliar question it falls back to keyword search, reads dozens of files, and burns its context window on incomplete answers. That is not an agent capability problem. It is a documentation discovery problem.

Read the white paper

Key takeaways

Inside your perimeter

Embeddings live in Delta. No data leaves Unity Catalog governance.

Hierarchical chunking

Document summaries first, precise chunks second. 42x token efficiency.

Scales to zero

Serverless SQL Warehouse computes similarity on demand, idle costs nothing.

$2 to $10 a month

Benchmarked against Vector Search and dedicated vector infrastructure.

About this white paper

Semantic search without the vector database

This paper documents how Covasant's engineering team solved documentation discovery for AI agents using only infrastructure already inside your Databricks workspace: Delta tables, Serverless SQL Warehouse, and the Foundation Model API. Full semantic search at $2 to $10 a month, with no dedicated vector database, no idle cost, and no data leaving your governance perimeter.

Written by the team that built and operates the platform
Includes the trade-offs we accepted, not only the wins

Read the architecture before you build it.

Twenty minutes of reading that can save a migration. No form, no gate, straight to the PDF.

Read the white paper Book a demo

The whitepaper identifies three failure modes. Vocabulary mismatch: when an agent asks about a concept using different terminology than the documentation uses, keyword search returns nothing. Context loss: keyword search returns individual file matches without preserving the relationships between documents, so agents lose the broader context needed for multi-step reasoning. Token economics: exhaustive file reading to compensate for poor search burns through context windows at a rate that is unsustainable for production AI agents doing repeated lookups.

Auraa stores document chunks and their embeddings in standard Delta tables within the existing Databricks workspace. At query time, the agent's query is embedded using the Foundation Model API and cosine similarity is computed on demand using Databricks Serverless SQL Warehouse. The warehouse scales to zero when not in use, so there is no idle cost. No data leaves the Unity Catalog governance perimeter. Total cost is $2-10/month at realistic usage levels.

Raw content chunks fail for conceptual queries because they lack document-level context. Hierarchical chunking generates a LLM-produced summary for each document alongside the raw chunks. When an agent queries, the system first matches against document summaries to identify the most relevant documents, then retrieves only the precise chunks from those documents. This two-stage approach means the agent receives focused, high-precision results rather than scanning all chunks. The 42x token efficiency advantage reflects the reduction in tokens consumed per successful answer versus exhaustive flat-chunk retrieval.

The whitepaper publishes a controlled benchmark evaluating all three approaches across 15 queries using three metrics: precision (did the result actually answer the query), token consumption (how many tokens did the agent use to reach the answer), and cost. The benchmark is designed to give teams the data to make an informed architecture choice rather than relying on vendor claims. The full results, including cases where simpler approaches were competitive, are published in the whitepaper.

The whitepaper publishes an honest cost comparison across three options: Delta tables with Serverless SQL Warehouse (Auraa's approach at $2-10/month), Databricks Vector Search (a managed vector index service with its own compute costs), and dedicated vector database infrastructure such as Pinecone or Weaviate. Annual projections are calculated at realistic usage levels for a production AI agent platform. The comparison includes idle costs, which are a significant factor given that most AI agent documentation queries are bursty rather than continuous.