This article introduces Robust-GAP, a hierarchical Retrieval-Augmented Generation (RAG) framework designed to eliminate semantic hallucinations and knowledge drift during multi-document log summarization. By combining dynamic causal graph extraction (DLCE), active topology verification (SGAV), and metadata provenance propagation (PAPP), the framework enforces strict citation traceability and prevents LLM-generated hallucinations.
Note: The complete academic preprint detailing this research is openly available on Zenodo (DOI: 10.5281/zenodo.21436390).
Standard Retrieval-Augmented Generation (RAG) pipelines fail when aggregating unstructured multi-document event streams because they lack causal tracking. When large language models (LLMs) summarize logs, chat transcripts, or transaction records, they frequently associate unrelated events. This behavior, known as knowledge drift, creates false dependencies.
For example, when troubleshooting an incident, an LLM might draw a causal link between a database table cleanup cron job and an unrelated disk write failure simply because both logs occurred within the same context window.
Robust-GAP resolves this issue. The pipeline constructs a dynamic causal graph of incoming events, verifies these connections against system schemas, and propagates source metadata directly to the final summary output as verifiable inline citations.
Robust-GAP is not an isolated development. It represents the fourth generation of an algorithmic lineage spanning a decade of research on hierarchical tree reduction:
Robust-GAP maps unstructured event streams into a two-level hierarchical merge tree. The pipeline relies on three core stages:
The following infographic illustrates the complete Robust-GAP workflow:
The table below contrasts Robust-GAP with alternative approaches:
| Feature / Metric | Vector + Batch | Hybrid + OneShot | Classic GAP | Robust-GAP (Ours) |
|---|---|---|---|---|
| Causal Topology | ||||
| No | No | Static Only | Dynamic (DLCE) | |
| Hallucination Pruning | ||||
| No | No | No | Yes (SGAV) | |
| Strict Provenance | ||||
| No | No | Textual Only | Metadata (PAPP) | |
| Citations | ||||
| 0% | 10% | 68% | 100% (Verifiable) | |
| Hallucination Rate | ||||
| High | High | Low | 0% |
To implement the pipeline, we released a Python-based CLI tool:
The script runs on Python 3 with zero external dependencies. It uses direct REST communication with the Gemini API, falling back to a local simulation mode if an API key is not configured.
The CLI handles two input JSON formats:
id
, timestamp
, text
, module
). Custom field keys can be mapped dynamically via flags like --key-text
or --key-module
.["error 1", "error 2"]
). The script automatically assigns sequential document IDs, timestamps, and schema wrappers on-the-fly.Execute the pipeline with:
python3 robust_gap_cli.py samples/travel_incident_data.json --language ja
data_file
: Path to the input JSON file containing log documents.--model
: Target Gemini model name (default: gemini-3.1-flash-lite
).--api-key
: API key for Gemini. Automatically loads from environment variables if omitted.--language
, -l
: Output language code (e.g. en
, ja
, zh
) (default: en
).--time-window
: Proximity window in seconds for DLCE (default: 1200
seconds).--cmdb-schema
: Comma-separated component lists for SGAV verification.--key-text
, --key-timestamp
, etc.: Dynamic key mapping for custom input structures.--mock
: Forces offline local mock simulation.We verified the pipeline against two pre-bundled datasets available in the samples directory using the live Gemini API.
This dataset (travel_incident_data.json) simulates travel disruptions mixed with unrelated noise logs (such as coffee purchases or postcard shopping) to test filtering.
python3 robust_gap_cli.py samples/travel_incident_data.json --language ja
Real LLM Output:
This dataset ([sre_incident_data.json](https://github.com/tanaikech/Robust-GAP/blob/main/samples/sre_incident_data.json)) simulates a multi-service outage spanning four parallel infrastructure dependency chains mixed with 16 background noise events.
python3 robust_gap_cli.py samples/sre_incident_data.json --language en
**Real LLM Output:**
During benchmarking on a 32-node cluster log dataset, we compared the hallucination and citation metrics across all four pipelines.
[Dxx]
), allowing instant programmatic verification of all claims.The core design of Robust-GAP—decoupled causal graphing, schema validation, and metadata propagation—applies to any domain requiring verifiable summaries of unstructured sequence logs.
When customer issues span multiple agents and chat channels, support logs get fragmented. Robust-GAP can ingest these unstructured conversations, isolate independent customer threads, and compile a chronological summary backed by direct message citations.
Financial audits require parsing thousands of transaction logs, ledger updates, and transfer records. By mapping transaction flows as a causal graph, Robust-GAP can flag suspicious transfer loops, validate them against compliance rules, and produce summaries of the anomaly with direct, ledger-level references.
For mission-critical systems, standard flat-vector RAG is insufficient. By enforcing structured causal hierarchies, Robust-GAP allows teams to deploy LLMs in production environments where hallucination risks must be strictly controlled, rather than just theoretically discouraged.
The codebase is open source and available on GitHub. Try running it on your own log streams.
If you use this work in your academic research, please cite the preprint as follows:
Tanaike, Kanshi. (2026). Robust-GAP: Hallucination-Resistant Hierarchical RAG with Dynamic Topology and Citation Provenance. Zenodo Report. https://doi.org/10.5281/zenodo.21436390