Unacog provides a web‑based platform that ingests large text collections, splits them into configurable token‑based chunks, and creates vector embeddings stored in a Pinecone index for semantic search. Users can experiment with chunk size, overlap, and two embedding pipelines—top‑K document aggregation or focused chunk‑plus‑context—to optimize retrieval relevance for downstream LLM applications.
Funding
Funding not disclosed
Founders
Product
Problem
Researchers and developers often need to perform semantic search over large, complex text collections such as academic papers or religious texts, but existing tools provide limited control over chunking granularity and retrieval strategies, leading to inconsistent search relevance and suboptimal context for downstream language models.
Solution
Unacog offers a web‑based library that ingests extensive text corpora, splits them into configurable token‑based chunks, and creates vector embeddings stored in a Pinecone index. Users can experiment with different chunk sizes and overlap settings to observe how granularity affects retrieval quality. The platform provides two embedding pipelines: one that aggregates top‑K chunks across multiple documents for broad context, and another that enriches a single matched chunk with surrounding text for deeper understanding. Results are returned as vectors that can be fed directly into LLMs, enabling more accurate semantic queries and generation. Interactive demos showcase the impact of these parameters on AI research papers and the English Bible, allowing users to fine‑tune their retrieval approach before integrating the service into their own applications.
Target Audience
Primary users are AI researchers, data scientists, and developers building LLM‑powered applications who need precise semantic retrieval over large text collections.
Features
- Configurable chunking with adjustable token length and overlap to optimize vector granularity
- Dual embedding pipelines: top‑K document chunk aggregation and focused chunk‑plus‑context embedding
- Pinecone‑backed vector index for scalable, low‑latency semantic search across millions of vectors
- Built‑in examples for AI research papers (2,800+ documents) and the English Bible with metadata tagging
- Real‑time query interface that visualizes how chunk size and retrieval method influence search results
- Compatibility with OpenAI’s text‑embedding‑3‑small model and other standard embedding APIs