emem-find-similar

This skill runs a nearest-neighbour search over the Tessera embedding field on emem.dev. Tessera is a 128-D learned multimodal vector that fuses Sentinel-2 optical, Sentinel-1 radar, and seasonality into one position-stable representation per cell per year. Two cells with cosine similarity >0.85 are usually the same physical archetype.

When to invoke

The user asks for analogues:

• "Find cities globally that look like Bangalore."
• "What other places have an urban canopy similar to Singapore?"
• "Show me regions with the same forest signature as the Western Ghats."
• "Compare Mumbai and Lagos by their Tessera embedding."

If the user wants exact-band matching (e.g., "all places with NDVI > 0.7"), this is the wrong skill — use query_region or compare_bands instead. This skill is vector cosine, not predicate filtering.

How to invoke

Step 1 — resolve the seed place to cell64

SEED_CELL=$(curl -sf -X POST https://emem.dev/v1/locate \
  -H 'content-type: application/json' \
  -d '{"q":"Bangalore, India"}' | jq -r '.cell64')
echo "seed cell: $SEED_CELL"

Step 2 — ensure the seed has a Tessera vector attested

/v1/find_similar returns 404 cid_not_found when the seed cell has no geotessera band attested on this responder. Materialise it first (idempotent if already present):

curl -sf -X POST https://emem.dev/v1/recall \
  -H 'content-type: application/json' \
  -d "{\"cell\":\"$SEED_CELL\",\"bands\":[\"geotessera\"]}" > /dev/null

Step 3 — query top-K neighbours

curl -sf -X POST https://emem.dev/v1/find_similar \
  -H 'content-type: application/json' \
  -d "{\"key\":\"$SEED_CELL\",\"k\":12}" \
  | jq '.neighbors[] | {cell, score, place: .place_label_cached, lat, lng}'

The response includes:

• neighbors[].cell — cell64 of the neighbour
• neighbors[].score — cosine similarity in [0, 1]
• neighbors[].lat, .lng — centre coords
• neighbors[].place_label_cached — cached human label if known
• neighbors[].band_used — almost always geotessera
• neighbors[].similarity_method — cosine (default) or hamming

(if you set band: "geotessera.bin128")

• neighbors[].deep_recall_url — the /v1/recall payload that

fetches the neighbour's full embedding for further drill-down

Picking the right vintage

geotessera defaults to the 2024 vintage. If the user asks "what looked like X in 2018?", you can change the band:

curl -sf -X POST https://emem.dev/v1/find_similar \
  -H 'content-type: application/json' \
  -d '{"key":"defi.zb493.xoso.zcb6a","k":12,"band":"geotessera.2018"}'

Available vintages: geotessera.{2017..2024} plus geotessera.multi_year (1024-D = 8×128 stacked, fuses all years). The multi-year vector picks up trajectory similarity — places that changed in similar ways.

Pitfalls

• **Cosine over Tessera reflects physical archetype, not socioeconomic

features.** A "similar" city may look the same from space (urban density, vegetation, water proximity) but be socially very different. Be honest with the user about what the similarity measures.

• Score thresholds are heuristic. >0.85 is "same archetype",

0.7–0.85 is "related", <0.7 is "weak". The actual cutoff depends on the seed cell's environment.

• First call may be slow if the seed needs materialisation; the

second is cache-warm.

Worked example

USER: Find cities globally that look most like Bangalore.

CLAUDE invokes this skill:
  POST /v1/locate {"q":"Bangalore"}    → defi.zb493.xoso.zcb6a
  POST /v1/recall {"cell":..., "bands":["geotessera"]}    → ok
  POST /v1/find_similar {"key":"defi.zb493.xoso.zcb6a","k":8}
    → neighbors:
       defi.zb5cf.nura.zd83c  0.91  New York City, USA
       defi.zb563.noxo.xAvu   0.89  Shanghai, China
       defi.zb541.…           0.87  Lagos, Nigeria
       …

CLAUDE replies: "By cosine over the Tessera 2024 embedding,
Bangalore's nearest neighbours globally are New York (0.91),
Shanghai (0.89), Lagos (0.87), and …. The similarity captures
shared physical archetype — dense urban core with patches of
green and surrounding peri-urban — not social or economic
features. Each result is content-addressed; click 'recall' on
any to fetch the underlying vector and verify it offline."