Highlighting and Snippet Extraction

When you receive query results from Vectara, alongside the result, you'll receive values for part_metadata and document_metadata. For example:

{
  "search_results": [
    {
      "text": "Foo bar baz.",
      "score": 0.2553684115409851,
      "part_metadata": {
        "lang": "eng",
        "section": "1",
        "offset": "36"
      },
      "document_metadata": {},
      "document_id": "doc_123456789",
      "request_corpora_index": 0
    }
  ]
}

For highlighting and snippet extraction, we need to pay attention to three key elements in this response:

1. text under search_results which gives us the text/length to highlight
2. section under part_metadata which tells us which specific section the relevant snippet showed up in
3. offest under part_metadata which tells us how far into the section the relevant snippet is

Text and offset

The most familiar parts of highlighting or snippet extraction if you've used any other search system are the text and offset values. The offset tells you how many characters into the given section should be skipped before any highlighting. Following that, the text tells you what the specific text is to highlight.

For example, if the original text was:

The quick brown fox jumped over the lazy dog. How vexingly quick daft zebras jump!

And if the query text is "striped horse-like animal," you might get back an an offset of 48 (how many characters before the sentence starting with "How" starts) and a text value of "How vexingly quick daft zebras jump!"

Configuring context

The context_configuration object in the query request allows you to control the amount of context included with each matching document part (snippet) that appears in a summary. Adding this context configuration affects the results quality for summarization by enhancing relevance and reducing ambiguity around each snippet. You can specify characters_before and characters_after or sentences_before and sentences_after to include before and after the snippet, as well as start_tag and end_tag that wrap the snippet, such as <b> and </b>.

"context_configuration": {
   "sentences_before": 2,
   "sentences_after": 2,
   "start_tag": "<b>",
   "end_tag": "</b>"
}

This example uses sentences_before and sentences_after. If you enter values for sentences before/after and characters before/after, then characters_before/after is ignored and summary returns the sentences values. Experiment and iteratve with different values.

Including additional context

Often, just having the text and offset values are enough to create a compelling highlighting/snippet extraction experience for short-form documents like social media posts and when you just have 1 section per documents. However, if you'd like to expand the snippet to include additional context, it's possible that the additional context may not be fully contained in the given section -- especially for longer documents. In those cases, you might want to include content from other sections.

If you created the sections yourself, you may choose to just replay the sectioning logic at query time and use as much of the additional sections as desired. However, if you aren't certain as to the section numbers and/or if you uploaded documents using the file upload API, then you might need to look up the additional sections. This can be done by an additional query to Vectara using filters. To do this, retrieve the document_id value of the document and perform a query for that ID. For example:

https://api.vectara.io/v2/query

POST https://api.vectara.io/v2/query
{
  "query": "",
  "search": {
    "corpora": [
      {
        "corpus_key": "my-corpus",
        "metadata_filter": "doc.id = '5b943498-d18c-4095-92f9-7a03f026f680'"
      }
    ],
    "offset": 0,
    "limit": 10
  }
}         

In this example, the relevant document ID is 5b943498-d18c-4095-92f9-7a03f026f680. We can then iterate through the results and show any number of sections either before (with a lower section number) or after (with a higher section number) as additional context for the snippet.