Semantic search

When the AI needs to check "Does the vendor have five years of experience in healthcare IT?", it doesn't read the entire 300-page proposal from start to finish. It uses semantic search to jump straight to the relevant passages.

How it works

After documents are parsed, the text gets split into small chunks (roughly paragraph-sized). Each chunk gets a mathematical fingerprint that captures its meaning - not just the words, but the concepts.

When the AI asks a question, that question also gets a fingerprint. The system compares fingerprints and returns the chunks whose meaning is closest to the question.

Why "semantic" matters

Traditional keyword search would look for exact words. If you search for "healthcare experience", it only finds passages containing those exact words.

Semantic search understands meaning. It would also find:

• "Our team has delivered electronic health records to three hospitals"
• "We completed a clinical data management project for St. Mary's Medical Center"
• "Medical sector projects make up 60% of our portfolio"

None of these contain the words "healthcare experience", but they're all relevant answers.

Why it's used

Two reasons:

Speed. Finding five relevant paragraphs in a 300-page document takes a fraction of a second. Having an AI read 300 pages takes minutes and costs much more.

Cost. Semantic search is cheap compared to AI processing. By narrowing down where to look first, the system saves significant AI costs. The AI only reads the passages that matter.

When the AI reads more

Semantic search is the first step. If the relevant passages don't fully answer the question, the AI can still request a broader read - "read pages 40-55" or "read the entire financial section." It's a practical balance between efficiency and thoroughness.

For complex requirements that span multiple sections, the AI might combine search results from different parts of the document to build a complete picture.