Core Capabilities

1. High-Precision Parsing

One entry for six formats, scanned documents readable, tables stay intact

The wide variety of enterprise document formats is the first hurdle in knowledge base construction. The platform ships with DeepDoc layout analysis (dual ONNX models for layout + table structure recognition, inferred locally), parses PDF, Word, Excel, PPT, OFD and XPS — six major formats — into structured data, preserving heading hierarchy, paragraphs, table row-column relationships (including merged cells) and reading order; scanned documents are recognized automatically via OCR.

2. Document Cleaning

Keep noise like headers, footers and watermarks outside the knowledge base

Headers, footers, page numbers, watermarks, duplicate content — these are background noise to humans but pollution to AI: page numbers suddenly appear in answers, and retrieval gets disturbed by duplicates. The platform automatically marks and cleans this noise after parsing, and fills back a cleaning report — what was processed and what was removed is fully documented.

3. AI Intelligent Redaction

13 categories of sensitive information, automatically masked before ingestion

Faces, ID cards, official seals, license plates, QR codes embedded in document images... letting them straight into the knowledge base is a serious leak risk. The platform reuses the compression engine's AI feature-protection detectors to automatically apply Gaussian blur/mosaic to sensitive regions before ingestion, and outputs a redaction report — what was detected in each image and where was processed is fully clear at audit time.

View the Full Data Security Framework →

4. High-Performance Compression

One pass, dual output: smaller files + cleaner knowledge

Compression is UGLYPEAR's core strength. Full coverage of 40+ formats across images, audio/video, PDF, Office, OFD, XPS and text, with an average compression ratio of 60% and up to 80%. Dual-metric SSIM/PSNR quality gating ensures smaller does not mean blurrier. For AI applications, image compression directly means lower multimodal inference token costs — measured savings of 85.9% on the GPT-4o basis.

Learn about the SmartSlim Compression Product Family →

5. Intelligent Chunking

Different documents, different cuts — semantics stay intact, retrieval stays precise

Fixed-length chunking is the industry's most common pitfall: a sentence is cut off mid-way, and AI retrieving half a piece of information can only guess. The platform selects a chunking strategy by document type and ships with four sets of preset parameters, fine-tunable per business:

6. Metadata & Lifecycle

Every knowledge point carries its own ID card, and outdated knowledge exits automatically

After chunking, every knowledge point is automatically bound to metadata such as document title, chapter path, publication time, version number and source department, and receives four-level content enrichment (breadcrumb navigation, keywords, summary, hypothetical questions). Document updates go through incremental processing: new versions enter the knowledge base automatically, old versions are automatically down-weighted and marked outdated, and content hashes ensure the same document is never ingested twice.

7. Permission Tagging

Who can see which knowledge, governed down to the knowledge-point level

Salary policies can be queried only by HR and management, and technical documents are isolated by department. The platform supports binding department, role, position, region and confidentiality-level labels to every knowledge point, and filters by permission before recall at retrieval time — permission checks happen at the data layer, not relying on the application layer's goodwill. The multi-tenant system naturally isolates knowledge bases of different customers/business units.

8. Three-Layer Evaluation

Let numbers speak for knowledge base quality, and hold bad answers accountable

A knowledge base is not finished once built. The platform ships with a three-layer evaluation system: corpus level (parsing completeness, table structure rate, OCR accuracy, redaction coverage), retrieval level (HitRate@K, Recall@K, MRR, NDCG@K), and generation level (faithfulness, answer relevance, context precision). After a user gives a thumbs-down, attribution runs automatically: whether the problem lies in chunking, cleaning, redaction or metadata can be found out in one check, and badcases flow back into the evaluation set automatically.

View Measured Data →