Research Intelligence System — Architecture

← Back
User Input Text · PDF · Image · URL · CSV / JSON ✎ Text Query 📄 PDF (PyMuPDF) 🖼 Image (Vision API) 🔗 URL (BeautifulSoup) 📊 CSV / JSON 0 · Intake Agent Detect type · Clean text · Extract topics · Classify intent OSS · No LLM · Zero token cost 1 · RAG / Knowledge Agent Keyword search in /knowledge/ · Identify gaps OSS · No LLM · Local file search 2 · Reasoning Agent Generate structured markdown answer · Summarise research LLM · claude-haiku-4-5 · Prompt cached 3 · Action Agent Extract metadata · Create or update .md note · Merge on conflict LLM · claude-haiku-4-5 · Prompt cached 4 · Linker Agent Inject [[wiki-links]] · Regenerate _index.md OSS · No LLM · Regex-based 5 · Validator Agent Score quality 1–10 · Flag broken links · Suggest improvements LLM · claude-haiku-4-5 · Prompt cached Output Answer · Quality Score (1–10) · Agent Trace · Input Type Badge Notes Created / Updated · Knowledge Base (sidebar) Git Sync (optional) In-memory cache Identical prompts skip API call entirely Anthropic prompt cache ~70% savings on repeated system prompt tokens
Open source · No LLM cost
LLM · claude-haiku-4-5 (or gpt-4o-mini)
Cost optimisation layer