genius-search-ingest

name	genius-search-ingest
skill_type	skill
description	Natively ingest Genius Agent web-search hits and crawled pages into the epistemic-graph knowledge graph — search results become :SearchQuery / :SearchResult / :SearchProvider / :WebPage typed nodes plus shared :Document text for semantic search. Use when the agent must persist what it searched or crawled into the KG so it becomes durable, queryable agent memory. Do NOT use just to view results (use genius-web-search) or to read a page body without saving it (use genius-web-crawl).
license	MIT
tags	["genius","knowledge-graph","ingestion","search","documents","epistemic-graph","mcp"]
metadata	{"author":"Genius","version":"0.1.0"}

Genius Search Ingest

Push the Genius search/crawl harvest into the ONE epistemic-graph knowledge graph as typed OWL nodes + documents, so a query and its evidence become durable, queryable agent memory instead of a throwaway result list.

When to use

• Persist a search's ranked hits into the KG (:SearchQuery + :SearchResult + :SearchProvider + :WebPage + :Document).
• Save crawled markdown pages as :WebPage + shared :Document text.
• Do both in one shot: run a live search and ingest it (genius_ingest_search).

When NOT to use

• Only display results to the user → genius-web-search.
• Only read a page body, no persistence → genius-web-crawl.
• Ingest arbitrary local files / a crawl --output-dir into graph-os generically → the platform mcp_graph-os_graph_ingest tool.

Prerequisites & environment

• A reachable epistemic-graph engine (graph-os). With no engine every entry point cleanly no-ops (returns None) — search/crawl still work.
• Ingestion rides the shared primitive agent_utilities.knowledge_graph.memory.native_ingest when present, else a self-contained txn over the lightweight GraphComputeEngine()._client.
• Provenance is stamped automatically: source=genius-agent, domain=genius.

Tools & actions

Python API in genius_agent.kg_ingest:

Function	Maps to
genius_ingest_search(query, max_results=10)	Wire-First: runs a real web search, then ingests it
ingest_search_results(query, results, provider=...)	Already-fetched hits → typed nodes + docs
ingest_crawled_pages(pages)	Crawled {url, markdown} pages → :WebPage + :Document
map_search_results(...)	Pure mapper → (entities, relationships, documents) (offline/testable)
ingest_entities / ingest_documents	Low-level typed-node / document writers

Node ids are genius:<class>:<externalId>; type matches the classes federated by genius_agent/ontology/genius.ttl.

Recipes

Search and ingest in one call (the Wire-First tool):

from genius_agent.kg_ingest import genius_ingest_search
genius_ingest_search("cilium eBPF service mesh", max_results=8)
# -> {"query": ..., "provider": "duckduckgo", "results": 8, "ingested": {"nodes": ..., "edges": ..., "documents": ...}}

Ingest hits you already fetched with genius-web-search:

from genius_agent.kg_ingest import ingest_search_results
ingest_search_results(
    "openraft consensus",
    [{"title": "openraft", "link": "https://github.com/...", "snippet": "..."}],
    provider="duckduckgo",
)

Ingest crawled pages from genius-web-crawl:

from genius_agent.kg_ingest import ingest_crawled_pages
ingest_crawled_pages([{"url": "https://docs.example.com/g", "markdown": "# Guide\n..."}])

Gotchas

• Everything is best-effort: no engine ⇒ None, never an exception.
• Result field names vary by provider; the mapper normalizes link/url/FirstURL and snippet/body/Text, but a hit with no URL is skipped.
• Retrieved text lands on the shared :Document class (never redefined by this package); hub-side enrichment chunks/embeds it for semantic search.
• Ids are content-hashed on (provider|query) and target URL, so re-ingesting the same query MERGEs (idempotent) rather than duplicating.

Related

• genius-web-search / genius-web-crawl — produce what this ingests.
• agent-utilities-source-integration — the fleet-wide native ingestion pattern.