Graph entity extraction produces fragmented, mistyped nodes and zero edges

[prakashUXtech]

A health check on a real personal soul ran soul repair --rebuild-graph. The rebuilt graph has 1881 nodes and zero edges. This is on current dev-level code with a fresh rebuild, so it is a code bug, not a stale graph.

What the rebuilt graph looks like

The nodes are not entities. They are single capitalized tokens, acronyms, and code identifiers lifted out of memory text. A sample:

• MERGED, SEPARATE, Literal, PRs, M3, M2a, Nov-2025, CI-green
• POCKETPAW_POCKET_REFRESH_SCHEDULER_ENABLED, CloudLifecycleHook, TaskSpec, HOME_POCKET_PROMPT

Nearly every node is typed person. Only a handful (pocketpaw, Claude, soul-protocol) come out as tool. With zero edges, graph traversal and graph-augmented recall have nothing to walk.

This is the same state #190 described before its planned overhaul: "64 entities, 0 edges, entity extraction is regex token-splitting on capitalised words, default type is person." #190 Phase 1 was supposed to replace that. The enum and provenance plumbing landed; the extractor did not.

Where the bug is

Extraction is regex token-splitting, not noun-phrase NER

MemoryManager.extract_entities() at src/soul_protocol/runtime/memory/manager.py:1893:

• Line 1906 splits the text into word tokens with re.findall(r"[\w][\w\-]*", combined).
• Lines 1916-1935 walk each sentence and, for every token after the first that starts with a capital letter and is not a stopword, create an entity. tokens[1:] is the whole sentence minus the first word, so MERGED, CI-green, TaskSpec, and POCKETPAW_POCKET_REFRESH_SCHEDULER_ENABLED all qualify. Compound names like "Soul Protocol" are split into two separate one-token entities.
• There is no noun-phrase chunking and no NER. The _TOPIC_PATTERNS pass only catches a few first-person speech templates ("I am learning X"), which do not fire on memory content written by the soul-sync hook.

Typing defaults to person

The capitalized-token branch hard-codes "type": "person" (manager.py:1932). The third-person relation pass does the same for any new subject or object (lines 1992, 2004). The only non-person path is the KNOWN_TECH lookup (lines 1908-1914), a fixed set of technology names. translate_to_ontology() (manager.py:311) runs afterward but only maps legacy synonym strings (technology to tool, user to person); it cannot infer a type. So anything not in KNOWN_TECH lands as person regardless of what it is.

The rebuild path drops relationships before they reach the graph

extract_entities does build edges in the third-person relation pass (manager.py:1973-2024) and returns them under each entity's relationships key. But repair --rebuild-graph discards them. src/soul_protocol/cli/main.py:3692-3698 builds the per-entity dict with only name and entity_type:

graph_ent = {
    "name": ent["name"],
    "entity_type": ent.get("type", "unknown"),
}

ent["relationships"] is never copied. update_graph() (manager.py:2063) reads entity.get("relationships", []), gets an empty list, and adds no edges. Even when extraction finds edges, the rebuild throws them away. That is the direct cause of zero edges after a rebuild.

Correct behavior

• Extract real noun-phrase entities. Keep compound names like "Soul Protocol" intact instead of shredding them into single tokens. Drop bare verbs, acronyms, and code identifiers, or route them to a concept type rather than person.
• Type entities through the ontology. A capitalized token is not evidence of a person. person should not be the catch-all default.
• Carry relationships through repair --rebuild-graph. Forward relationships (and edge_metadata for provenance) into the dict passed to update_graph() so a rebuilt graph has edges.

References

• feat(graph): tier overhaul — typed ontology, real extraction, graph API #190 — graph tier overhaul (typed ontology, real extraction). Phase 1 set the bar at "<5% type errors, typed entities, non-empty provenance." The current rebuild does not meet it.
• graph: entity extraction skipped for low-significance interactions causes graph plateau #220 — related graph-extraction bug (significance gate skipping extraction).