Research: Search and Entity Resolution

Scope: forge.find text / vector / hybrid recall for half-remembered entities.

Executive summary

Exact-match Cypher fails when an LLM emits "alice c." instead of "alice-chen". forge.find is the v0.5.0 answer for lexical, vector, and hybrid lookup over node properties.

Key findings:

Text search handles natural-language descriptions, paraphrases, stems, and case variants. It does not reliably handle single-character typos or abbreviations.
Search latency stays low at notebook scale; index build cost dominates at 10K+ nodes, not query time.
Index name and description (and aliases) together — description text is what makes paraphrase recall work.
Hybrid search needs real embeddings to beat text-only ranking; mock vectors do not demonstrate semantic quality.
CamelCase tool names recall better when descriptions are indexed; store acronyms explicitly in aliases when needed.

Pattern

from graphforge import GraphForge

forge = GraphForge("entities/")

# After bulk ingest, rely on forge.find over indexed properties
table = forge.find("software engineer acme", label="Entity", limit=5)

# Hybrid when you already have query vectors
table = forge.find(
    "alice engineer",
    label="Entity",
    vector=query_vec,
    limit=10,
)

Query style	Expectation
Exact / partial name tokens	Strong
Description paraphrase	Strong if description indexed
Stems (`engineering` → engineer)	Strong
Typos / abbreviations	Weak — use aliases or preprocessing
Meaning without keyword overlap	Needs vectors + hybrid

Recommendations

Always index descriptive text with names.
Prefer batch indexing after bulk MERGE ingest; incremental updates for live graphs.
Use forge.find as the primary entity-resolution entry point in LLM and agent flows.