Context Graph Agent

The context graph agent runs after curated records are available. It builds a sparse graph of useful relationships between decisions, constraints, facts, preferences, references, evidence, and handoffs.

The pipeline below shows the DSPy module flow.

---
config:
  flowchart:
    curve: linear
---
graph TD;
    __start__([<p>__start__</p>]):::first
    load_inventory(load_inventory)
    build_candidates(build_candidates)
    link_records(link_records)
    review_links(review_links)
    persist_graph(persist_graph)
    __end__([<p>__end__</p>]):::last
    __start__ --> load_inventory;
    load_inventory --> build_candidates;
    build_candidates --> link_records;
    link_records --> review_links;
    review_links --> persist_graph;
    persist_graph --> __end__;
    classDef default fill:#f2f0ff,line-height:1.2
    classDef first fill-opacity:0
    classDef last fill:#bfb6fc

Inputs

• active durable records for one project
• semantic-neighbor candidate pairs
• existing graph edges for duplicate avoidance

Flow

1. load_inventory loads active durable records and existing graph edges.
2. build_candidates builds semantic-neighbor clusters and candidate record pairs.
3. link_records asks DSPy to propose sparse, grounded relationships.
4. review_links asks DSPy to drop weak, duplicate, or generic links.
5. persist_graph writes graph nodes, graph edges, and semantic cluster labels.

Clustering

The persisted graph stores one durable cluster layer:

• semantic clusters from semantic-neighbor records

A planned hosted dashboard can derive Louvain communities and combined visual lenses from accepted graph links without adding transient visualization labels to the local runtime store.

Output

The graph projection is derived context. Durable records stay canonical. context_nodes and context_edges are refreshed from curated records and are kept ready for clustered graph exploration in the planned hosted product.