Doclime

Performs vector-based semantic search over uploaded PDF documents and academic papers by converting natural language queries into embeddings and matching them against indexed document embeddings. Uses dense retrieval (likely transformer-based embeddings like BERT or specialized academic models) rather than keyword/BM25 matching, enabling the system to understand research intent and find conceptually related papers even when keyword overlap is minimal. The indexing pipeline processes PDFs on upload, extracting text and generating embeddings that are stored in a vector database for fast approximate nearest neighbor retrieval.

Enables users to ask natural language questions about specific PDF documents and receive extracted answers without manual reading. The system likely uses a retrieval-augmented generation (RAG) pipeline: when a user queries a document, the system retrieves relevant text chunks from the PDF using semantic similarity, then passes those chunks to an LLM to generate a contextual answer. This combines document chunking (splitting PDFs into overlapping sections), embedding-based retrieval, and LLM inference to provide document-specific answers with source citations.

Allows users to query across multiple uploaded PDFs simultaneously to synthesize findings, identify contradictions, or compare methodologies across papers. The system likely uses a hierarchical RAG approach: retrieving relevant chunks from each document based on the query, then using an LLM to synthesize or compare the retrieved information. This requires managing context across multiple documents, deduplicating similar findings, and generating comparative summaries that highlight agreements and disagreements across sources.

Processes uploaded PDF files to extract text content and prepare it for semantic search and querying. The system handles PDF parsing (converting binary PDF format to text), text cleaning (removing headers, footers, page numbers), and chunking (splitting text into overlapping segments for retrieval). The extracted and chunked text is then embedded using a transformer-based embedding model and stored in a vector database for fast retrieval. This pipeline must handle diverse PDF formats, including scanned documents (via OCR if supported) and complex layouts.

Automatically expands or reformulates user queries to improve semantic search results by understanding research intent. When a user enters a query like 'machine learning for medical diagnosis', the system may expand it to include related terms like 'deep learning', 'clinical decision support', 'diagnostic AI', and 'neural networks for healthcare' before performing retrieval. This likely uses query expansion techniques such as synonym injection, semantic paraphrasing via LLMs, or learned query reformulation models. The expanded queries are then used to retrieve more relevant documents from the vector database.

Implements usage-based access controls on the freemium tier, capping the number of documents users can upload, queries they can perform, and API calls they can make. This is a business model enforcement mechanism that limits free users to a subset of platform capabilities (estimated <100 documents, <50 queries/month) while offering unlimited access on paid tiers. The system tracks usage per user account and enforces limits at the API level, returning rate-limit errors when users exceed their quota.

Automatically extracts structured metadata from uploaded PDFs, including title, authors, publication date, abstract, and keywords. This likely uses a combination of PDF header parsing (extracting text from the first page) and NLP-based named entity recognition (NER) to identify author names and publication dates. The extracted metadata is stored alongside the document embeddings and used for filtering search results, displaying document information, and organizing the user's document library. This enables users to see paper details without opening the full PDF.

Uses a vector database (likely Pinecone, Weaviate, or Milvus) to store and retrieve document embeddings at scale. When a user uploads a PDF, the system chunks the text, generates embeddings for each chunk using a transformer model, and stores the embeddings in the vector database with metadata (document ID, chunk index, text preview). During search, the user's query is embedded using the same model, and approximate nearest neighbor (ANN) search is performed to retrieve the most similar chunks. This architecture enables fast semantic search even with thousands of documents and millions of chunks.