Modules

GCTRL ships four cooperating modules that form a single pipeline from raw sources to agent-ready memory: KEX (extraction), FUSE (fusion / entity resolution), Manage KGs (lifecycle), and Talk-to-Graph (retrieval). Each is reachable from the portal UI and over the API/MCP layer. Endpoint paths below are illustrative.

KEX - Knowledge Extraction

KEX is the ingestion front door. It turns raw, unstructured input into structured graph fragments plus embedded text chunks.

What it does

• Ingests raw text and files - documents, plain text, and uploads.
• Parses code repositories deterministically - repo structure, symbols, and relationships are extracted by rule rather than guesswork, so the same repo always yields the same graph.
• Runs NER + relation extraction - identifies entities and the relationships between them.
• Chunks and embeds text - splits source text into passages and writes embeddings into the vector store for later semantic retrieval.
• Tags classification + source provenance - every extracted item carries its classification (clearance) and a pointer back to its exact source.

Execution model

KEX is job-queue based. Ingestion runs as asynchronous jobs so large corpora process without blocking, and job status is queryable.

Endpoints (illustrative)

FUSE - Knowledge Fusion / Entity Resolution

FUSE merges many extraction jobs into one clean, unified knowledge graph. Without it, repeated ingests produce a pile of near-duplicate entities; FUSE collapses them into single canonical nodes.

Entity resolution (high level)

FUSE uses a multi-stage resolver:

1. An exact / near-exact pre-filter quickly matches obvious duplicates.
2. A semantic resolver with blocking compares remaining candidates within efficient blocks rather than across the whole graph.
3. An optional graph-embedding link-prediction stage surfaces non-obvious links the earlier stages miss.

The resolver's tuning is intentionally not published here - this is a high-level description of the stages only.

Beyond merging

• Builds HOT dossiers - compiled authoritative summaries for important entities.
• Detects communities and central nodes - finds clusters and the high-influence nodes within them.
• Distills human-readable Wiki pages - turns the graph into curated prose.

Endpoints (illustrative)

Manage KGs

Manage KGs is the lifecycle layer for knowledge compilations. A compilation can be a RAW graph (the structured entities/relations) or a WIKI view (the curated prose surface).

Capabilities

• CRUD compilations - create, read, update, and delete RAW graphs and WIKI views.
• Graph traversal - search entities, fetch neighbors, compute shortest paths.
• Provenance / lineage - trace any fact back to its source and ingestion job.
• Corrections - fix wrong facts so they are never re-extracted; the correction sticks across future ingests.
• Audit - every change is logged.
• Export - extract compilations for downstream use.
• Scheduled refresh - keep compilations current on a cron schedule (incremental new-only or full regeneration).

Talk-to-Graph (RAG)

Talk-to-Graph is the retrieval and question-answering layer over everything KEX and FUSE built.

Retrieval

• Hybrid retrieval - combines vector (semantic) and lexical (keyword) search over text chunks and the knowledge graph, so structural facts and fuzzy recall arrive together.
• Clearance-filtered - results are scoped to what the caller is allowed to see.

Answering modes

• Single-pass - fast, one retrieval round for direct questions.
• Agentic deep mode - multi-hop reasoning that traverses the graph across several steps for complex questions.
• Citations - answers point back to the chunks and entities they were built from.

Privacy modes

• GDPR incognito - sessions live in browser memory only; nothing is persisted server-side.
• Standard mode - conventional session handling for non-sensitive workflows.

See also

Architecture · Memory Layers · Agents & MCP