Finding Anchors¶

Anchors are the structurally important nodes relative to your seeds, found with Personalized PageRank. They answer where to look before any traversal happens, which makes them useful both as a standalone orientation tool and as a way to sharpen a retrieval.

Calling it¶

Give find_anchors() your seeds and how many nodes you want back:

anchors = engine.find_anchors(seeds, topn=20)

Argument	Default	Meaning
`seeds`	required	Seed entities for personalized PageRank
`topn`	`20`	How many top nodes to return

It returns (node_id, ppr_score) tuples sorted by importance, highest first:

anchors = engine.find_anchors(seeds=["community/insurance_claims"], topn=20)
for node_id, ppr_score in anchors[:10]:
    print(f"{ppr_score:.4f}  {node_id}")

topn is the main dial. A small value (5 to 10) gives sharp focus on the few most important nodes; the default of 20 is a balanced neighborhood view; a large value (50 or more) maps a wider region at the cost of more work downstream.

Anchor-then-retrieve¶

The most valuable pattern is a two-step flow: use anchors to discover the important nodes around a broad seed, then retrieve paths from only the strongest of them.

# 1. Find the important nodes around a broad seed
anchors = engine.find_anchors(seeds=["community/insurance_claims"], topn=10)

# 2. Retrieve paths from the strongest anchors
strong = [node_id for node_id, score in anchors[:3]]
result = engine.retrieve(seeds=strong, max_paths=50, hop_limit=3)

This narrows a vague starting point down to the parts of the graph that actually matter before you spend any exploration budget, which is often the difference between a noisy retrieval and a sharp one. The same idea shows up in Tuning Retrieval as the single biggest lever on quality.

Inspecting one node¶

When you want to study a single node rather than rank many, get_neighbors() returns its immediate neighborhood with relation types and directions:

info = engine.get_neighbors("entity/provider_456")
print(info["degree"])
for n in info["neighbors"][:10]:
    print(n["direction"], n["rel"], n["id"], n["weight"])

Next¶

The concept behind all of this is Personalized PageRank, and Your First Retrieval shows where anchors fit in the full pipeline.