Skip to content

RAG search

RAG search is the heart of Myeline: ask a natural-language question, get a synthesised answer with source citations.

Pipeline

User question
  ↓
bge-m3 embedding (local Ollama)
  ↓
ChromaDB hybrid search (RRF + MMR over the selected scopes)
  ↓
HyDE / multi-query / contextual retrieval (auxiliary LLM)
  ↓
Cross-encoder reranking (auxiliary LLM)
  ↓
Synthesis (local Ollama OR external BYOK API depending on edition)
  ↓
Answer + citations [1], [2], [3]…

Embedding is always local, regardless of edition. Synthesis:

  • In pure sovereign: local Ollama (Mistral-Nemo, Llama 3.1, Mixtral depending on what you host)
  • In sovereign-hybrid: local Ollama by default, or per-organisation switch to Mistral / Claude / OpenAI / Gemini with your key (BYOK)

Citations

Every answer cites its sources as clickable [1], [2]. Each source displays:

  • Document title or article title
  • Author / source when available
  • Date when available
  • Excerpt (~500-token chunk) that contributed to the answer
  • Relevance score

If an answer has no citation, the LLM didn't find support in the documents — either the base lacks the information, or the question is too vague.

Strict mode

Toggle in the UI: when enabled, the LLM is forced to answer only based on retrieved sources (explicit system prompt + post-validation). If no source covers the question, the answer is: "I couldn't find the information in the available sources."

Recommended for:

  • Regulatory / legal research
  • Medical research
  • Any context where a hallucination would be a risk

Single-document

From the library, clicking a document opens a chat scoped to that single document. Useful for:

  • Querying a long report ("what are the recommendations?")
  • Comparing sections ("do the conclusions of part 3 contradict part 1?")
  • Extracting numerical data ("what is the 2024 revenue mentioned?")

Writing tips

  • Precise questions > vague ones. "What are the explicit-consent exceptions in GDPR art. 6?" works better than "tell me about GDPR".
  • Context: state the domain if your sources cover several topics ("in French employment law, …").
  • Iterate: refine the question across multi-turns (see Conversations) rather than dumping everything at once.

Known limits

  • Documents > 200 pages: chunking can dilute links between distant sections. Prefer single-document chat with contextual retrieval enabled.
  • Complex tables: imperfect PDF extraction (limits of pdfplumber/docx parsers). Prefer a CSV export when possible.
  • Mixed languages in the same base: bge-m3 is multilingual (100+ languages), but retrieval quality degrades when FR and EN mix in the same question.