Retrieval

Every query starts with a prune step (namespace, folder, collection, and as-of date), then runs one or more retrieval lanes over what is left, and always returns path citations you can hand straight to a reader. From any hit, an agent can navigate outward into the surrounding document or the entities it mentions.

Prune (namespace, folder, collection, as-of)→Search, then navigate→Path citations

Hybrid searchFull-text keyword search, vector similarity, and 1-hop REL expansion from those seeds, fused with RRFEntry point for direct questions and keyword-shaped queries

NavigateOpen a section, hop to a parent, sibling, neighbouring chunk, or a related entity, from a hit you already haveReading outward in long structured documents, and following typed relationships

You choose which search lanes run, alone or together. Each fails independently: a soft failure in one lane empties that lane's results without blocking the others.

Hybrid search

Hybrid search fuses three lanes over the same Falkor graph: full-text keyword search on the lexical overlay, vector similarity on the vector overlay, and 1-hop relationship expansion on the domain overlay, combined with Reciprocal Rank Fusion. Full-text and vector run together; relationship expansion then walks from those seed chunks. Hybrid search only returns chunks that are enabled (or that have no enabled stamp yet). Disabled namespaces, folders, and files stay in KG Studio, grayed out, and do not appear in retrieval. This is the entry point for direct questions and keyword-shaped queries, and it works the moment a namespace reaches searchable_embedded.

Outline navigation

The structural overlay is a navigable table of contents. kg_outline returns a file’s sections in document order, each with its section key, title, depth, parent, and page range, paged for long files and limitable by depth.

Document root:OutlineNode

1. SectionCHILD_OF root

Chunk 12Chunk 13

2. SectionCHILD_OF root

2.1 Subsectionopened by section key

Chunk 27

Citationnamespace/path.md > 2. Section > 2.1 Subsection (pp. 4-5)

Reading the outline costs titles, depths, and page ranges, not full chunk text. An agent picks a section key from it, opens that section's covering chunks, and hops to the parent, siblings, or children from there.

An agent opens a section’s covering chunks with kg_open on that section key, and moves around the tree with kg_hop: outline_parent, outline_siblings, and outline_child. prev_chunk and next_chunk step through the document in reading order. Every chunk that comes back carries its own section context (ancestors, siblings, and children) and a full path citation (namespace/path.md › Section › Subsection, pages 4-5), so an agent that lands mid-document through search can read outward without starting a new query. See MCP for the tool arguments and caps.

Entity navigation

Entity navigation starts from a named entity, found by name match, vector similarity, or both, optionally filtered to published entity types. From there an agent hops along typed relationships one step at a time (mentions, rel, same_type), and every hit is a chunk package carrying the evidence quote behind the edge. An entity stays visible only while at least one enabled, readable chunk still mentions it. This is what KG Studio’s domain graph view uses for “show me what connects to ESMA,” and agents reach it as a direct tool call.

Path citations

Every retrieval call returns citations shaped as namespace/path › Section › Subsection, pages, not a bare chunk id. This means an agent’s answer, or a KG Studio result, always has a human-readable trail back to the exact document section a claim came from, and clicking a citation opens the Reader view at that location.

Collection and as-of scoping

Retrieval accepts a collection filter and an as-of date. This keeps regulated, versioned corpora from mixing unless you explicitly ask for both: a query scoped to mifid-3 never silently blends in mifid-2 chunks, and an as-of query only sees relationships that were valid at that point in time. See Collections and temporal for the full model.

Soft failure per lane

Each search lane fails independently. If the vector overlay has a transient issue, hybrid search still returns full-text and relationship-expansion results rather than an empty response, and the failure is logged for your platform’s observability so you can see it happened without losing the rest of the answer.

Automatic context per agent

You set how much Knowledge Graph context an agent gets without asking for it, per agent, in AI → Agents.

No namespacesNo Knowledge Graph context on any turn, and no retrieval toolsNarrow task agents that work only on their inputs

Tools onlyNamespaces assigned and the Knowledge capability on; the agent retrieves when it decides toAgents that need the corpus occasionally

Auto-search onHybrid search on each new user prompt injects matching chunk packages before the turnAssistants answering from a governed corpus

You set these per agent in AI → Agents: the Knowledge Graph namespaces tile, the Knowledge capability tile, and the Auto-search knowledge graph on new prompt toggle, which is off until you turn it on.

• Namespaces and tools: assign corpora on the Knowledge Graph namespaces tile and enable the Knowledge capability so the agent can call retrieval tools at all. An agent with neither gets no Knowledge Graph context.
• Ambient search: Auto-search knowledge graph on new prompt runs a hybrid search over the assigned namespaces on each new user prompt and injects the matching chunk packages before the turn starts. It is off until you turn it on, so a narrow task agent pays no latency or cost for context it will not use.
• Threshold: the agent’s KB injection score threshold decides which of those matches are placed in context, with per-thread overrides in chat. A chat control turns Knowledge Graph search off for a single conversation.