How our RAG Pipeline is Different

The UBIK RAG (Retrieval-Augmented Generation) pipeline is engineered to solve the fundamental performance bottlenecks that plague standard implementations. Rather than relying on generic, off-the-shelf components, we have built a system that optimizes every stage of the retrieval process—starting with how information is extracted from your documents.

​

Bottleneck #1: Parsing & Extraction

The first and most critical step in any RAG pipeline is Parsing: extracting usable text and structure from raw files. Many basic libraries attempt to parse documents quickly but fail to preserve structural information or miss content entirely. A fast parser might extract the text but lose the context of headers, tables, or layout, leading to fragmented and confusing chunks for the LLM. If you feed bad content into your pipeline, you will never retrieve good answers. UBIK solves this by providing adaptive parsing pipelines tailored to the document type, required speed, and the depth of information needed. We support a wide range of formats including standard documents (PDF, DOCX, CSV, Excel, JSON, Text), web content, and multimodal files (MP4, MP3). Once a file is processed, we apply the appropriate parsing strategy based on the document’s characteristics and your user preferences.

​

Our Parsing Pipelines

We offer three distinct pipelines to balance speed, cost, and extraction quality:

​

1. Low Latency Pipeline

• Best for: High-volume processing where speed is critical and documents are simple text.
• Mechanism: Quickly extracts raw text without deep layout analysis.
• Trade-off: It is the fastest option but the least robust. It may skip complex layout elements or fail to extract text from image-heavy documents.

​

2. Standard Pipeline (Layout-Aware)

• Best for: Classic business documents (PDFs, DOCX, CSVs) where structure matters.
• Mechanism: Leverages OCR (Optical Character Recognition) and Computer Vision to detect and preserve the document’s layout.
• Customization:
  • Default: Uses the platform’s default OCR engine (Mistral OCR).
  • Optimized: You can deploy our proprietary, optimized OCR model in your own admin section for enhanced performance and privacy needs. Reach out to us via email at contact@ubik-agent.com for a demo on spinning up a custom instance.

​

3. Visual / Improved Pipeline (Multimodal)

• Best for: Complex documents with charts, images, or non-textual information.
• Mechanism: Converts documents into visual representations (PDF/Image) and applies a Vision Language Model (VLM) to “read” the document like a human would.
• Capabilities: This pipeline can extract meaning from images without text and process video content to create rich representations.

​

Specialized Format Handling

For specific modalities that require unique processing, we leverage dedicated parsers:

• Audio/Video (MP3, MP4): Transcribed and processed to extract both spoken content and visual context.
• Websites: Scraped and cleaned to remove boilerplate while preserving article structure.
• Code Files: Parsed to retain syntax highlighting and structural indentation.

By using a set of proprietary processing techniques, UBIK ensures that we extract not just the raw text, but the structural and multimodal context of your information. This high-quality extraction is the foundation of a high-performance RAG system.

​

Bottleneck #2: Encoding & Representation

Once you have extracted high-quality content, the next challenge is Encoding transforming that information into a format that a machine can understand and retrieve effectively. Most standard RAG pipelines rely on a Single Index approach. They convert all your text into a single vector representation using one embedding model. While simple, this approach often fails to capture the nuances of complex documents. A single vector might struggle to represent both the semantic meaning of a paragraph and the visual context of an accompanying chart (most of the time this gets ignored because multimodality in RAG is hard to make).

​

The Multi-Signal Approach

UBIK overcomes this limitation by allowing you to mix and match multiple signals to create a richer, more granular representation of your information. Instead of relying on a single embedding, you can leverage different models to capture various aspects of your data:

• Textual Semantics: Use a model optimized for understanding the core meaning of the text.
• Multilingual Capabilities: Incorporate a model specifically trained to handle multiple languages, ensuring accurate retrieval across global content.
• Visual Context: Integrate embeddings that represent the visual elements extracted during the parsing phase (e.g., charts, diagrams, images).

By combining these signals, you create a Multi-Vector Representation for each document chunk. This allows the retrieval system to discriminate between pieces of information with much greater precision. For a deeper dive into our philosophy on this topic, read our glossary entry on Multi-Signal Search.

​

Bottleneck #3: Retrieval & Search Strategy

Even with perfect extraction and encoding, a RAG pipeline can fail if the Retrieval mechanism is too simplistic. Standard systems typically rely solely on Cosine Similarity or dot product calculations against a single semantic index. While effective for broad conceptual matching, this approach has two major flaws:

1. False Positives: It often retrieves content that is semantically “close” in vector space but factually unrelated to the specific query.
2. Keyword Blindness: Pure semantic search can miss documents that contain the exact keywords you are looking for (e.g., specific product codes, error IDs, or proper names) because the embedding model focuses on general meaning rather than specific terms.

​

Hybrid Search: The Best of Both Worlds

UBIK addresses this by implementing a robust Hybrid Search engine. We don’t just look for meaning; we look for exact matches and combine them intelligently.

• Semantic Retrieval: Leverages the multi-signal vectors (text, visual, multilingual) discussed above to find conceptually relevant information.
• Keyword Matching: Simultaneously runs keyword-based algorithms (like BM25) to identify documents containing the exact terms from your query.

By fusing these two approaches, we ensure that if you search for “Error 505 in Module X”, we find documents that discuss “Module X” errors (semantic) and specifically mention “505” (keyword), ranking them higher than a general article about “System Errors”.

​

Customizable Fusion & Weighting

The “magic” lies in how these different signals are combined. UBIK gives you control over the Fusion Algorithm and Signal Weights directly from your User Preferences. You can configure:

• Fusion Algorithms: Choose how results from different indexes are merged (e.g., Reciprocal Rank Fusion).
• Index Weights: Assign higher importance to specific signals. For example, you might weight the “Keyword” signal higher for technical documentation or the “Visual” signal higher for design assets.

​

Bottleneck #4: Precision & Reranking

Even after a sophisticated hybrid search, the initial set of retrieved results may still contain noise. You might get 50 “relevant” chunks, but only 5 of them actually contain the answer to your specific question. This is where the Reranker comes in. A reranker is a specialized system that takes your query and the candidate results, analyzes them deeply, and assigns a Relevance Score to each one. It acts as a strict filter, discarding irrelevant information and re-ordering the rest so the most meaningful content is processed by the LLM.

​

Reranking Strategies

UBIK offers a spectrum of reranking options to balance speed, cost, and intelligence:

• API-Based Rerankers: (e.g., Jina, Cohere) Leverage smaller, optimized models for extremely fast scoring. Great for high-volume applications.
• LLM-Based Rerankers: Use a Large Language Model to reason about the relationship between the query and the document chunk. This provides much higher accuracy for complex queries.
• Vision-Language Rerankers (VLM): This is what makes a pipeline fully multimodal. By using a model that can “see,” we can rerank charts, images, and video frames based on their visual content, not just their text descriptions.

​

Advanced Configuration & Security

In your user preferences, you will see a list of available rerankers tailored to the models currently active on the platform. This selection includes:

• llm_tool_calling Binary Rerankers: Specialized models that output a binary (relevant/irrelevant) decision or a precise score using function calling capabilities.
• llm_multimodal Rerankers: Advanced models (like GPT-4o, Claude 4 Sonnet, Gemini 1.5 Pro) capable of analyzing both text and images simultaneously for the highest possible precision.
• API Rerankers: Optimized external services like Jina or Cohere.

Self-Hosting for Maximum Security For use cases requiring maximum security and data isolation, you are not limited to the platform’s default models. You can self-host your own reranking model and register it for your specific user account. If you need to deploy a custom self-hosted model, please reach out to us via email at contact@ubik-agent.com. We will provide a demonstration and guide on how to spin up an instance and integrate it. By configuring the right reranker, you ensure that your agent only sees the information that truly matters, reducing hallucinations and improving answer quality.

​

Bottleneck #5: Generation & Citation

The final stage of the pipeline is synthesizing the retrieved and filtered information into a coherent answer. Once the relevant information has been filtered by the reranker, it is passed to a Generative Model to construct the final response.

​

Model Selection & Flexibility

You can select the exact model architecture that fits your needs—whether it’s a classic high-speed model, a “thinking” model for complex reasoning, or a specialized domain model. Key Capabilities:

• Model Selection: Choose the best model for your specific use case (e.g., GPT-4o for reasoning, Claude 3.5 Sonnet for coding/writing).
• Multimodal Synthesis: If you select a vision-compatible model, the pipeline leverages full multimodality, allowing the model to “see” and interpret the visual extracts (charts, diagrams) preserved from your documents.
• Precise Citations & Highlighting: This step closes the loop by leveraging the structural data preserved during the initial Parsing phase. Because we maintained the document’s layout and context from the start, the model can accurately quote the exact source and highlight meaningful extracts for the user. For audio and video content, this includes specific timestamps, ensuring full traceability across all modalities—whether it’s a page number in a PDF or a specific minute in a meeting recording.

​

Scaling for Performance

Ultimately, this design empowers you to scale your RAG pipeline according to your specific constraints: optimize for ultra-low latency with faster models, or prioritize maximum performance and reasoning depth with larger ones. This flexibility allows you to choose the best tool for the job.