SureLink AI discovers and indexes text from your local files and feeds the most relevant excerpts into a large language model (LLM) during a query. This is a retrieval‑augmented generation (RAG) workflow designed for answers with citations, not a full‑text replacement for industry-specific tools.
1) Content & file‑handling limits
- Text extractability matters. SureLink AI performs best with machine‑readable text. Scanned PDFs, images, handwriting, and diagrams need OCR and/or Image description turned; without it, the system will only document the meta-data of the file.
- Formatting fidelity is lossy. Complex tables, formulas, pivot sheets, tracked changes, comments, and footnotes may lose structure when converted to plain text, which can reduce retrieval precision.
- Unsupported or restricted files. Password‑protected/encrypted files, very large compressed archives, or files with embedded binaries may be skipped by the indexer.
- Links & attachments. Embedded links or referenced files outside the indexed scope won’t be followed unless those targets are also linked and indexed.
- Versioning & staleness. Updated files are not retroactively reflected until re‑synced; deleted files may persist in caches until the next scheduled sync.
What this means for you: Prefer original, text‑native sources; run OCR on scans; run Image Recognition on pictures.
2) Retrieval & answer‑quality limits
- Data Link Limitations: SureLink AI can reliably link up to 100,000 files. Quality of responses may vary after this limit.
- Token budgets force summarization. SureLink AI can’t send “all your docs” to the model; it selects and compresses top passages. Important context can be missed if it’s not retrieved.
- Conflicting sources. When two documents disagree, the LLM may blend or hedge. Ask for direct quotes with file and page to anchor answers.
- Multilingual & jargon. Cross‑language retrieval and domain‑specific acronyms depend on your embedding model; expect lower recall unless your corpus and embeddings are aligned.
- Determinism. LLM outputs vary run‑to‑run. For regulated uses, refer to citations and enable human review before decisions.
3) Hard technical limits
- Prompt limiting: SureLink AI limits users to 400 prompts per day
- Image Recognition limiting: SureLink AI can run image recognition up to 5,000 images
- Template limiting: 10 templates, and 100 tags per template
- Model in use: SureLink AI currently supports retrieval augmentation via gpt-oss-120B. Supported input context window: Up to 65,536 tokens in total, which includes both your retrieved document snippets and your user prompt combined. Model output limit: Up to 16,384 tokens per response.
- Rate limiting: Groq enforces organization‑level limits by requests/tokens per minute and per day. When exceeded, you’ll receive HTTP 429 with headers (e.g., x-ratelimit-*, retry-after)—your client must back off and retry.
4) Scale & operations limits
- Index size & throughput: Practical limits depend on your storage/index choice and hardware. Very large corpora increase index time and can raise tail latencies for retrieval.
- Change detection: Frequent file edits can trigger re‑chunking and re‑embedding; plan for scheduled re-syncs.
- Access control at answer‑time: If user permissions change after indexing, filtered retrieval may still surface titles or metadata unless you Clear Links.
5) Customer‑side best practices (to stay within limits)
- Turn on OCR when linking PDF’s where you want to query based on the file content
- Turn on Image Scanning for images where you want the system to generate a text summary of the file.
