fmSR AI v1

Slide 1Seven AI Patterns

Vienna is calling. Answer the damn phone

Matt Navarre · navarre.ai · free classes every Thursday

Slide 2Why short text breaks embeddings

Not enough signal

Two or three words gives the model almost nothing to embed. A 40-word paragraph produces a rich, specific vector. “Bob Smith” produces a vague one – it ends up near every other two-token name.

Surface-form variance

Smith / Smyth / Smithe / Смит. Robert / Bob / Bobby / Rob. Embeddings care about meaning, not about orthography, and there isn't much “meaning” in a proper noun.

Cultural & transliteration drift

Accents. Non-Latin scripts. “Müller” vs “Mueller.” “李娜” vs “Li Na.” These are the same person to a human and nearly unrelated to a vector.

Slide 3Give GenAI 5 seconds. Get a richer record.

STAGE 3

STORED NAME

GENAI

EXPANDED TEXT

EMBED → STORE

Bob Smith

Generate variants

• Bob Smith
• Robert Smith
• Smith, Robert
• Bobby / Rob
• Smyth / Smithe

[0.13, -0.44, 0.88 ...]

SYSTEM PROMPT · person

“Generate realistic name variants for use in a semantic search index. Always include: the original name, LastName-FirstName form, accent-free spelling, common nicknames or diminutives, transliteration variants if relevant. Return a plain text list – one variant per line.”

Everything goes into the embedding. The record side is now fat with signal.

Slide 4Now do the same on the query. Many-to-many.

STAGE 4

ENRICHED RECORD

ENRICHED QUERY

• Bob Smith
• Robert Smith
• Smith, Robert
• Bobby / Rob / Robb
• Smyth / Smithe
• ABC Technology
• bob@abctech.com

• Robert Smyth
• Rob Smyth
• R. Smyth
• Smyth, Robert
• robertt@abc-tech.com
• 237 More Park Dr, CA
• Alphabet Tek?

MATCH

Orthogonal variants meet in the middle. Typos, nicknames and transliterations fall on the same point in vector space.

Slide 5Same pattern, more context per record

ENTITY → COMPANIES

WHAT WE EMBED

REAL-WORLD MATCH

User types:

• ABC Technology
• 237 More Park Drive, California

• Original org name
• Abbreviations & acronyms
• With / without legal suffix
• Address – full and normalized
• List of staff names
• List of staff emails

Finds:

• Alphabet TekSystem
• 2337 Moorpark, Santa Monica CA

✓ Prevents duplicate data entry.

Slide 6Project names? Same playbook.

ENTITY → PROJECTS

PROJECT NAME → VARIANTS

• Financial Audit Preparation Law Office
• Monetary Audit Preparation Lawyer
• Prepare for review of financial documentation
• Accounting document analysis – Law Office

Your user typed “prep for law office audit.” They'll find it.

The script only enriches the project NAME field. That's deliberate – and leads straight into the next slide about what we do NOT embed.

Slide 7Semantically clever. Numerically blind.

THE HONEST LIMIT

Meaning, similarity, paraphrase

“law office audit” matches “Financial Audit Preparation Law Office” because the WORDS mean similar things.

Dates, amounts, ranges, exact strings

“Projects in Q2” vs a due-date of 2026-05-14 – a vector will not know those are the same thing. Numbers are not meanings.

Right tool rule:

semantic search for meaning + a plain WHERE clause for dates and amounts.

Stop trying to make one tool do both. That's the #1 mistake I see with new FileMaker AI work.

Slide 8Re-enrich and re-embed on a schedule

EXTRA CREDIT

Projects, products and companies change. So do their embeddings.

• Scheduled
• server script

• Re-embed &
• store

Re-enrich

Nightly / weekly

Fresh variants

Insert Embedding

New staff? New price? New project description? The server re-enriches the text and re-embeds. Search stays accurate without a human touching anything.

(The Update Embedded Record script already does this – aim a Schedule at it.)

Slide 9Teach the index what “cheap” means

ENTITY → PRODUCTS

Users search “cheap red shoes.” The database stores $39.99 / crimson / Nike.

PRICE TIERS

COLOR FAMILIES

CATEGORY SYNONYMS

• Shoes:
• < $50 → cheap, budget
• $50–150 → mid-range
• > $200 → luxury, premium
• Cars:
• < $20k → basic, economy
• > $50k → luxury, flagship

• Red →
• crimson, cherry,
• cardinal, burgundy,
• wine, garnet, brick,
• tomato, rose, maroon
• Silver → platinum,
• ash, pewter, titanium

• Shoes →
• sneakers, trainers,
• kicks, running shoes
• SUV →
• crossover, 4x4,
• sport utility

“Cheap” is not a price. It's a category-relative judgment. Bake that in at write-time.

Slide 10The cost story: basically free

~100

small

tokens in

tokens out

model OK

the search string + a compact system prompt

5–10 name or query variants – that's it

gpt-4o-mini, haiku, local models. No frontier needed.

In practice: false positives are rare.

Real-world noise (typos, nicknames, abbreviations) is what we were after. The enrichment surfaces genuine overlap – not random neighbors.

Slide 11Where a CS purist will push back

“Use hybrid search”

Production systems often combine BM25 (lexical) + semantic + structured filters, merged with rank fusion. Our Stage 3+4 is a strong subset – adding a keyword pass for SKUs is cheap and worth doing when you have exact-token fields.

“Your index goes stale”

When data changes and you don't re-enrich, search accuracy drifts. That's why the scheduled re-embed matters. Build it in from day one.

“Price-tier text is baked in”

If pricing strategy shifts (“now $150 is mid-range for shoes”), the existing embeddings encode the old worldview. Re-enrich the catalog. Not the end of the world – just a scheduled job.

All of these are solvable. None of them invalidate the approach.

Slide 12QUESTIONS?

• Build the search
• your users think they have.

REACH FOR IT WHEN

DON'T REACH FOR IT WHEN

• You store short-text entities
• Users type sloppy searches
• Duplicate prevention matters

• You need exact-match by ID
• Data is numeric or date-based
• A WHERE clause does the job

navarre.training · free classes every week · matt@navarre.training