Knowledge-Guided Context Completion (KGCC)

• KGCC is a framework that integrates retrieval-based summarization with conditional generation to fill gaps in structured knowledge systems.
• The methodology serializes three steps—summarization, gap-finding, and guided generation—to produce precise, contextually relevant evidence.
• In medical QA, KGCC improves accuracy by reducing noise from retrievals, achieving benchmark gains of 12.5% and 4.5% over traditional methods.

Knowledge-Guided Context Completion (KGCC) is an advanced framework for imputing or inferring missing elements in knowledge-based systems (such as knowledge graphs or domain-specific corpora) by leveraging structured, unstructured, and parametric sources of knowledge to generate or retrieve contextually relevant supplemental information. The objective of KGCC is to identify the gap in current available evidence, direct the system to produce or retrieve complementary background knowledge, and integrate this knowledge for robust downstream reasoning or prediction, as exemplified in the MedRGAG framework for medical question answering (Li et al., 21 Oct 2025).

1. Conceptual Foundations and Formal Objectives

KGCC addresses intrinsic limitations of both retrieval- and generation-based approaches to knowledge-intensive tasks. In retrieval-augmented generation (RAG), external corpus search may yield incomplete, irrelevant, or noisy documents; generation-augmented generation (GAG) relies solely on a model's parametric memory and risks hallucinations or factual errors.

The central objective of KGCC is to "complete the context" by:

• Summarizing retrieved evidence to distill only knowledge directly useful for a given task (e.g., medical QA).
• Explicitly identifying knowledge points or semantic components that remain missing after retrieval.
• Conditionally generating background documents or explanations that address these knowledge deficits.
• Integrating both retrieved and generated evidence for answer or inference.

Let $Q$ denote the task (e.g., a question), $D_{ret}$ the set of retrieved documents, and $D_{gen}$ the set of generated documents. KGCC formalizes a completion operator $\mathcal{C}$ such that the finalized evidence set is $\mathcal{C}(Q, D_{ret}, D_{gen})$, where $D_{gen}$ is produced by guiding the generator according to the knowledge gaps found in $D_{ret}$ (Li et al., 21 Oct 2025).

2. Methodological Framework

KGCC is typically implemented as a modular sequence of three steps interleaved between retrieval and downstream prediction:

1. Summarization of Retrieved Knowledge: A strong LLM is instructed—with task-specific prompts—to condense each retrieved document $d_i \in D_{ret}$ into a set of essential knowledge points relevant to $Q$. Non-informative or spurious information is explicitly disregarded, for instance via outputs such as “No useful information” for irrelevant passages.
2. Exploration of Missing Knowledge: An explorer module, given $Q$ and the set of relevant knowledge summaries $\{s_i\}$, is prompted to produce a structured enumeration of the key knowledge points not present in $D_{ret}$ but critical for a comprehensive response. This step defines the set of missing knowledge $\mathcal{K} = \{k_1, ..., k_m\}$, which may correspond to facts, concepts, or explanatory statements.

$$\mathcal{K} = \mathcal{M}_e(Q, \{s_i\}; \mathcal{P}_e)$$

where $\mathcal{M}_e$ is the explorer model and $\mathcal{P}_e$ is the prompt template.

1. Conditional Generation of Complementary Context: For each $k_j \in \mathcal{K}$, a generator model (e.g., an LLM) is prompted with $Q$ and $k_j$ to generate focused background documents $d_j^{gen}$. If $|\mathcal{K}|$ is less than the target number of context documents, additional generations are conditioned only on $Q$.

$d_j^{gen} = \mathcal{M}_g(Q, k_j; \mathcal{P}_g)$

for $j=1,...,|\mathcal{K}|$, where $\mathcal{M}_g$ is the generative model and $\mathcal{P}_g$ is the generation prompt.

All generated and retrieved documents are then aggregated for downstream answer synthesis or prediction (Li et al., 21 Oct 2025).

3. Integration with External and Parametric Knowledge

Unlike approaches that solely trust retrievals (RAG) or unconstrained generation (GAG), KGCC unifies external and parametric knowledge in a synergistic manner:

• External Knowledge: Provides factuality, grounding, and verifiability. KGCC filters and condenses this information, extracting only what's directly relevant per query.
• Parametric (Model) Knowledge: Used exclusively to fill in identified gaps, and is steered via explicit prompts about missing knowledge, reducing the likelihood of hallucination and off-target content.

The overall pipeline can be formalized as follows (see Algorithm 1 of (Li et al., 21 Oct 2025)):

1. For each retrieved $d_i$, summarize as $s_i$ via LLM with summarization prompt.
2. For $Q$ and $\{s_i\}$, extract $\mathcal{K}$, the set of missing knowledge.
3. For each $k_j \in \mathcal{K}$, generate $d_j^{gen}$ using a conditioned prompt.
4. If $|\mathcal{K}| < k$ (desired number of background docs), sample additional $d_j^{gen}$ conditioned only on $Q$.
5. The union $D_{ret}^{summ} \cup D_{gen}$ forms the evidence set for downstream answer generation.

4. Impact on System Performance and Reliability

In the MedRGAG system, integration of a KGCC module led to substantial improvements in multiple medical QA benchmarks: a 12.5% improvement over MedRAG (retrieval-only) and a 4.5% gain over MedGENIE (generation-only) (Li et al., 21 Oct 2025). Ablation studies confirm that removing KGCC measurably reduces answer accuracy and reliability.

Key effects:

• Context is ensured to be both comprehensive (as in, all key knowledge points are covered) and precise (low noise, minimal irrelevant information).
• Hallucinations in generated evidence are reduced, as the generator’s output is directly anchored in discovered gaps.
• Final answer or inference is constructed with a minimal, non-redundant, and maximally informative evidence set.

A plausible implication is that for any downstream application requiring justification or verifiable inference, KGCC offers robustness against both overfitting to parametric knowledge and undercoverage from external retrieval.

5. Comparison with Other Completion Approaches

KGCC represents a distinct paradigm relative to:

• Standard RAG: Only retrieves documents, lacking an explicit mechanism to fill knowledge gaps. Tends to produce incomplete answers when retrieval is noisy or misses critical evidence.
• GAG: Generates all evidence based on model parameters, highly flexible but vulnerable to hallucinations and confabulation. KGCC constrains generation using retrieval-derived missing knowledge signals.
• Retrieval+Unconstrained Generation: May try to supplement retrieval with arbitrary LLM outputs, but without systematically determining what knowledge must be generated, potentially producing non-relevant expansions.

In contrast, KGCC serializes (summarize $\rightarrow$ gap-find $\rightarrow$ guided generate) the process, ensuring that only targeted, contextually appropriate knowledge points supplement the retrieved set (Li et al., 21 Oct 2025).

6. Real-World Applications and Generalizations

While demonstrated in the medical QA setting, KGCC as an architectural concept is transferable to any structured inference or reasoning task where:

• External retrievals are likely noisy or incomplete.
• The generative model alone cannot guarantee coverage or factuality.
• Domain-specific constraints require that evidence be both comprehensive and justified.

Potential application domains include regulatory, financial, legal, or scientific question answering; context completion in scientific knowledge graphs; and automated decision support in highly specialized fields.

The core modular structure—summarization, gap-finding, and conditional generation—serves as a blueprint for extending KGCC to other domains, models, or integration strategies.

7. Summary Table: KGCC Pipeline in MedRGAG

Step	Function	Core Output
Summarization	Extract only useful knowledge from retrieval	Set of concise, task-relevant summaries
Exploration	Identify critical missing knowledge points	Structured set $\mathcal{K}$ of knowledge gaps
Conditional Generation	Generate context per identified gap	New documents $d_j^{gen}$ addressing gaps
Evidence Integration	Aggregate summaries and generated docs	Comprehensive, non-redundant evidence set

This workflow ensures comprehensive, reliable, and context-completed evidence for downstream knowledge-intensive reasoning (Li et al., 21 Oct 2025).