Metadata Enrichment Strategies
Metadata is more than labels. It is signal amplification.
Course outline & contents
1. Metadata > Vectors
Vectors are fuzzy. Metadata is precise. Sometimes, the best search isn't semantic—it's structural. Enriching your chunks with extrat data improves filtering and ranking.
2. Enrichment Type A: Contextual Summary
If you split a document, Chunk #52 doesn't know it belongs to a "Legal Contract." Strategy: Prepend the document summary to the chunk.
Original: "The tenant must pay by the 5th."
Enriched: "(Context: Lease Agreement 2024, Section: Payments) The tenant must pay by the 5th."
Now, the vector "knows" this is about a lease payment, not a dinner bill.
3. Enrichment Type B: Questions Generation
Users ask questions. Documents contain statements. There is a "Mismatch." Strategy: Ask an LLM to generate 3 questions that this chunk answers, and embed those questions.
- Chunk: "The API limit is 500 requests."
- Enriched: "What is the rate limit? How many calls can I make? The API limit is 500 requests."
Now, a user asking "rate limit" hits a direct semantic match.
4. Summary
Don't just index what is written. Index what is meant. Metadata enrichment bridges the gap between the user's intent and the author's text.
Key Intuition: "Embed the question, not just the answer."