Adding Attributes

A step-by-step scenario: how to add a new attribute to an existing anchor (or link).

See Attributes vs Link. In short:

  • if the value is scalar and used only for description → attribute;
  • if the value references an entity that has its own attributes or other relationships → link.

2. Make sure the data exists

An attribute is the schema for a value that already lives in the data. If the source table doesn’t have the corresponding column, the attribute won’t work (the LLM will see the description but Cypher won’t return values).

Check that:

  • the Grist Data doc or your source table has the right column;
  • values are non-empty and the format matches the expected dtype.

In Grist > Data Model > Anchor_attributes:

attribute_nameanchordescriptiondata_exampledtypeembeddableembed_thresholdquery
descriptionproductDetailed product description for users.”MacBook Air, M2 chip…“strtrue0.65MATCH (p:product {product_id: $id}) RETURN p.description
weight_kgproductProduct weight in kg.1.24floatfalse—MATCH (p:product {product_id: $id}) RETURN p.weight_kg
release_dateproductRelease date.”2024-06-10”datefalse—MATCH (p:product {product_id: $id}) RETURN p.release_date

For an edge attribute, do the same in Link_attributes, putting the edge sentence in the link column.

4. Pick embeddable

Data typeembeddable?
name, title✅
description✅
address✅ (carefully — comparisons may beat exact matches)
FAQ.question✅
any numeric❌
date / time❌
boolean❌
code, SKU, article❌
URL, file path❌
enum (structured set)usually ❌

5. Pick embed_threshold

For embeddable attributes only. Use the canonical starting table from Tuning Embeddings & Thresholds:

Attribute kindStart
Names / exact identifiers (product, employee, brand)0.75–0.85
Descriptions and explanatory text0.65–0.75
Long documents / chunks0.50–0.65
FAQ entries0.70–0.78

After running on the golden dataset, adjust. Too many false positives → raise; missed valid matches → lower. Full tuning loop and edge cases (long-vs-short text, common terms across anchors) — see Tuning Embeddings & Thresholds.

6. Run ETL

Backoffice → ETL → Run Selected on memgraph_steps (or do a full run). Datapipe is incremental: generate_embeddings is keyed by (node_id, node_type) for anchors and (from_node_id, to_node_id, edge_label) for edges. When a row in nodes/edges or in dm_anchor_attributes/dm_link_attributes changes (including flipping embeddable or editing embed_threshold/attribute_name), Datapipe re-runs the step only for the affected keys. A bulk change like adding an embeddable attribute on an anchor will recompute every node of that type once; subsequent runs are cheap.

7. Verify

In Memgraph Lab:

MATCH (p:product) RETURN p.description LIMIT 5

Values for the new attribute should be returned.

In chat, ask a question that requires the new attribute:

“Which product weighs less than a kilogram?”

If it works — great. If not — see Details:

  • is the attribute visible in the system prompt? (no → ETL didn’t apply / description is empty)
  • was the Cypher generated correctly? (no → improve description and the data_example)

Checklist

  • The data column exists and is populated.
  • dtype matches the actual format.
  • embeddable is set correctly (only for text).
  • embed_threshold is set (if embeddable).
  • query works — verified in Memgraph Lab.
  • ETL ran without errors.
  • A test question gives the correct answer.

Common mistakes

  • dtype doesn’t match the data. ETL fails or writes incorrectly.
  • embeddable=true for numbers/SKUs. Useless — embedding “999.00” carries no semantics.
  • Empty query. The assistant knows the attribute exists but can’t fetch it.
  • Description is too generic. “Property of product” — the LLM won’t know when to use it.
  • One attribute “does it all”. Don’t pack JSON into one column if queries on it matter — split into multiple attributes.

What’s next