Skip to content

Context Embedding Research Brief

The questions behind one idea: attach situating context (source and event summaries) to claims, the way contextual RAG attaches document summaries to chunks, so that matching can reason beyond exact hard-claim hits. Created 2026-09-01. Updated 2026-09-01. Reports produced from this brief carry a 90-day TTL.

Framing (priors to test, not premises)

From Agus's proposal (2026-09-01), stated as hypotheses:

  1. The RAG analogy holds. Prepending a generated summary of the whole document to each chunk measurably improves retrieval; attaching source and event summaries to claims should analogously improve matching, because a claim's meaning depends on where it came from.
  2. Context unlocks LLM-judge matching. With claim provenance context available, an LLM can decide whether a partial entity belongs to an existing entity in cases hard identifying claims cannot settle.
  3. Context is derived data. Summaries should attach through lineage (claims already point to their partial, event, and sources), not by widening the Claim schema.

Researchers must report disconfirming evidence with the same weight as confirming evidence.

Q1. Mechanics of contextual retrieval in RAG

How is chunk-context generation actually done (prompt, model, context length, cost), and what are the measured gains? What variants exist (contextual embeddings, late chunking, RAPTOR, parent-document retrieval) and when does each win?

Unblocks: whether the technique is real and transferable, and which variant maps to Indra.

Q2. Transfer from retrieval to entity matching

Retrieval failure and match failure are different problems. Does the entity-resolution literature support LLM judgment with contextual evidence for match decisions, and at what accuracy and cost?

Unblocks: where in the pipeline an LLM judge may sit, and what it is allowed to decide.

Q3. Where and when to generate summaries

Per source, per event, or both? At ingestion, at composition, or lazily at match time? Which model class, and what does it cost at Helmi volume?

Unblocks: whether summary generation enters v1 ingestion scope, and an amendment to the "deterministic core" principle (summarization would be a second LLM stage).

Q4. Risks and guardrails

Hallucinated or skewed context, invariant erosion (context-based judgment must not mint member links), and how to eval the judge.

Unblocks: the guardrail lines in the PRD if the idea is adopted.