Automatic Literature Survey Tool

Keeping up with scientific literature became impossible

In the R&D sector, the rate of publication has reached an inflection point. For researchers and analysts, staying current with the "state of the art" is no longer a human-scale problem. Traditional keyword search is inefficient - returning thousands of irrelevant results or missing crucial papers due to vocabulary mismatch. The cognitive load of filtering, reading, and synthesizing this vast corpus is a significant drain on high-value intellectual capital.

Semantic understanding at scale

Building a tool that can truly "read" and "understand" thousands of documents required overcoming several NLP challenges:

• Polysemy: The same word means completely different things across fields - "model" in machine learning versus in architecture or fashion. Context-aware disambiguation was essential.
• Structured extraction: The tool needed to extract and aggregate specific findings - methodologies, results, benchmarks - across multiple heterogeneous sources and present them comparably.
• Domain generalization: The system had to work across diverse terminologies - from biomedical research to financial news - without requiring per-domain retraining.

A RAG-based knowledge engine

TendersLab developed the Automatic Literature Survey Tool, a Retrieval-Augmented Generation system combining vector search precision with LLM reasoning capability.

1. Vector embedding and semantic search

Documents are ingested into a vector database (Pinecone/Milvus) by converting text chunks into high-dimensional embeddings using OpenAI's text-embedding-3. This enables semantic search - finding documents conceptually similar to a query, not just lexically matching. A query for "novel battery materials" retrieves papers discussing "solid-state electrolytes" even if the exact keywords differ.

2. LLM-driven extraction and summarization

Once relevant documents are retrieved, a chain of LLM prompts performs specific extraction tasks. The system can answer complex questions like "What were the accuracy metrics for all ResNet variants mentioned in these papers?" by synthesizing data points from multiple PDFs into a structured comparison table - with citations.

3. Dynamic knowledge graph construction

Beyond Q&A, the tool constructs a lightweight knowledge graph mapping relationships between authors, institutions, and key concepts. This lets researchers visualize the "genealogy" of an idea and identify influential papers or emerging research clusters - connections that a keyword search would never surface.

A force multiplier for research teams

• Weeks to hours: Literature surveys that previously took a team weeks can now be generated in hours, complete with citations and executive summaries.
• High citation accuracy: By grounding outputs in retrieved document chunks (RAG architecture), fabrication risk is significantly reduced - teams trust the output.
• Cross-domain discovery: Researchers routinely discover relevant methodologies from adjacent fields they would have otherwise never encountered.