Tacit Knowledge Retrieval System
Hybrid retrieval over laboratory discussions combining RGT embeddings with a knowledge graph, built end-to-end from document ingestion to evaluation.
- Retrieval
- Knowledge Graph
- LLM Systems
- Evaluation
A research internship project for preserving and retrieving the tacit knowledge embedded in unstructured laboratory discussions.
Problem
Laboratory knowledge is often buried inside unstructured discussions — meeting notes, chat logs, and informal write-ups. When that knowledge is needed again, people search manually or ask around, and much of it simply gets lost.
Research Question
Can a hybrid retrieval system that combines embedding-based search with a knowledge graph improve the discovery of tacit knowledge compared with dense retrieval alone?
System
The pipeline runs end-to-end: raw discussion data in, retrievable answers out.
Discussion documents
│
▼
Document ingestion ── normalize, chunk, dedupe
│
├──────────────► RGT embeddings ───────────┐
│ ├─► Hybrid Retrieval
└──────────────► Knowledge Graph ──────────┘ │
(rules + NLP) │
▼
Reranking
│
▼
Evaluation
- RGT embeddings capture semantic similarity between passages.
- Knowledge graph captures entity and relationship structure that embeddings alone miss.
- Hybrid retrieval fuses both signals with reranking before evaluation.
My Contribution
- Designed the document ingestion and chunking strategy for noisy discussion logs.
- Built the embedding generation and graph construction stages.
- Implemented the hybrid retrieval and reranking logic.
- Ran the evaluation suite and error analysis.
Experiments
- Baselines: dense retrieval only, BM25, and graph-only traversal.
- Proposed: hybrid (dense + graph) with reranking.
- Evaluated on a labeled query set derived from real lab discussions.
Results
Hybrid retrieval improved recall on relationship-heavy queries, where pure dense retrieval struggled. Error analysis showed the graph mainly helps when the answer is spread across multiple documents.
Results are placeholder text — replace with actual numbers and analysis.
What I Learned
- Retrieval quality is decided as much by ingestion and chunking as by the model.
- Knowledge graphs and embeddings are complementary, not competing, signals.
- Evaluation needs a task-aligned query set, not just off-the-shelf benchmarks.