-
Notifications
You must be signed in to change notification settings - Fork 1
Bridge Entity
type: concept up: "Markov-Retrieval-Tutorial" tags: [graph, corpus-structure, retrieval, vocabulary] related: ["Vanilla-RAG-Failure-Mode", "Heterogeneous-Graph", "Tutorial-Corpus"]
A bridge entity is one whose chunks use vocabulary from more than one cluster, so that the random walk can cross clusters by passing through it.
Embedding spaces cluster vocabulary. Canonical-vocabulary entities sit in one region, tail-vocabulary entities in another, and a query can typically only reach one region by direct similarity. A bridge entity is one that mentions concepts from both regions in the same chunk. Its chunk-to-concept edges connect the two regions in the heterogeneous graph even though their concepts are far apart in embedding space.
A walker starting at query concepts in one region can reach the other region by:
- Stepping from a query concept to a chunk that mentions both vocabularies (the bridge chunk).
- Stepping from that chunk to a concept in the canonical region.
- Stepping from that canonical concept to other canonical chunks.
Without the bridge, no short path exists, and the chain cannot manufacture one.
In the tutorial corpus, HoosierMetals is the only fully bilingual entity. Its chunks use canonical defense vocabulary ("defense primes," "ITAR registered," "AS9100") and procurement vocabulary ("DLA," "Defense Logistics Agency," "MIL-spec") in the same passages. For the DLA query, this is what makes recovery possible at all.
Figure 2 of the tutorial PDF shows the recovery path for the DLA query: from the query concepts, through HoosierMetals, into the canonical defense cluster.
The DLA query also exposes a clear limit: HoosierMetals is the only bridge in the corpus, and one bridge is not enough to populate the connecting region with sufficient signal. The chain recovers HoosierMetals but cannot bridge into the broader canonical cluster (where ITAMCO, AcmeMfg, and AeroForge live) with mass that competes for top-five ranking.
The path-length decomposition in Figure 3 makes this concrete: HoosierMetals' overview chunk has a clean k=1 spike (total mass 0.0315) while ITAMCO's capabilities chunk has flat, small contributions across all path lengths (total mass 0.013).
See Path-Length Decomposition for what those numbers mean and Tutorial Queries for the broader query results.
In a real corpus, bridges arise naturally from people who write in both vocabularies (operations staff using procurement jargon while describing canonical capabilities). In a synthetic corpus, you have to put them there. The tutorial corpus has exactly one because the goal is to show what one bridge can and cannot do.
This is also where the extraction-retrieval co-design point lands: an extractor that canonicalized "DLA" to "Defense Logistics Agency," or that tagged procurement and defense-prime concepts as related at indexing time, would have built the bridge into the graph instead of leaving the chain to discover it through co-occurrence.