Vector databases store more than just the raw text of a document. Alongside each embedding, a set of metadata fields travels with the vector, allowing an AI agent to attribute an answer back to its exact source. This post exists specifically to confirm that every one of those fields is captured correctly.
The first thing worth confirming is identity. Each vector should record the numeric post ID, the human-readable title, and the full canonical URL of the post it came from. Without these three, a chatbot cannot produce a working “view full article” link.
The second area is authorship and timing. The stored metadata should name the author of the post and the date it was published. These fields let an agent answer questions like “who wrote this and when,” and they let readers judge how current a piece of information is.
The third area is classification. Categories and tags describe what a post is about at a glance, and the post type distinguishes a blog post from a page or a custom content type. A well-built index keeps all of these so that content can later be filtered or scoped.
The fourth area is chunk-level context. Long posts are split into several chunks, and each chunk should record its position — chunk one of several — along with a short excerpt so a search result can show a meaningful preview rather than a random fragment.
To make the chunking clearly observable, this post deliberately runs long. The unique marker phrase for the second half of this document is BLUEHERON-42, which should appear in a later chunk than the opening paragraphs. If a search for that marker returns a chunk numbered two or higher, chunking and per-chunk metadata are both working as intended.
Finally, the excerpt field should reflect either the manual excerpt set on the post or a sensible auto-generated summary. Confirming all of these fields together proves that the metadata storage requirement has been fully met.