RAG Pipeline

QARK’s RAG (Retrieval-Augmented Generation) pipeline turns your documents into a searchable knowledge base. Drag in files, they get parsed, chunked, and embedded into a local vector index. Ask questions with @document-search and get answers with inline citations pointing to exact sources, pages, and relevance scores.

Everything processes locally through your configured providers. No cloud upload.

---

Smart routing

Not every document needs a full vector pipeline. QARK routes based on size:

Document size	Route	What happens
Below threshold	Direct injection	Full text inserted into the prompt context. No chunking, no embedding. Status: direct
Above threshold	Full RAG pipeline	Parsed → chunked → embedded → indexed → searched at query time

The threshold is a percentage of the model’s context window (default: 30%). A 200-page PDF always goes through RAG. A 2-paragraph markdown file gets injected directly.

Configure in per-conversation RAG settings or globally in Settings → RAG → Direct Injection Threshold.

---

Supported file types

Format	Extensions	Notes
PDF	.pdf	Text extraction + optional image extraction
Word	.docx	Preserves heading structure
Excel	.xlsx	Sheet-aware parsing
PowerPoint	.pptx	Slide-by-slide extraction
Markdown	.md, .mdx	Heading-based chunking
HTML	.html, .htm	Tag-aware parsing
Plain text	.txt, .log, .csv	Line-based chunking
EPUB	.epub	Chapter-aware extraction

---

Add documents

Three methods:

1. Drag and drop — drag files or folders onto the conversation. Multiple files accepted simultaneously.
2. File picker — click the attachment button in the composer.
3. Recursive folder scanning — drop a folder. QARK scans recursively, preserving directory structure in the document panel.

---

Processing pipeline

Each document progresses through visible stages. The document panel shows real-time status per file:

Stage	Description
pending	Queued for processing
parsing	Extracting text from the file format
chunking	Splitting into semantically meaningful segments
embedding	Generating vector embeddings for each chunk
indexing	Storing vectors in the local search index
searching	Being queried (visible during response generation)
ready	Indexed and available for retrieval
error	Processing failed — hover for details
skipped	Unsupported file type or empty file
direct	Small enough for direct injection — no RAG processing

---

Search strategies

QARK selects a retrieval strategy per query. Four are available:

Semantic search

Standard vector similarity. Your query is embedded and compared against chunk vectors using cosine similarity. Best when your query language closely matches the document language.

HyDE (Hypothetical Document Embedding)

Generates a hypothetical answer first, embeds that, then searches for similar chunks. Closes the vocabulary gap — your query “What’s the refund policy?” matches chunks describing the policy without using the word “refund.”

Step-back prompting

Generates a broader, more abstract version of your query, then searches with both original and abstracted queries. “Why did Q3 revenue drop in EMEA?” becomes “What factors influenced EMEA regional revenue trends?”

Auto (default)

QARK analyzes the query and picks the best strategy automatically. The selected strategy appears in the response metadata.

---

Code-aware chunking

For source code files, QARK uses syntax-aware chunking:

• Functions, methods, and classes kept as whole units
• Code blocks in Markdown never split mid-block
• Import statements and module headers stay attached to the first chunk

Retrieved code chunks are syntactically complete, not arbitrary text fragments cut at a character limit.

---

Cross-encoder reranking

Reranking is optional but improves retrieval accuracy. The process:

1. Over-fetch — vector search retrieves 3x the final chunk count as candidates
2. Re-score — a cross-encoder reranking model (Voyage AI or Jina AI) scores each candidate against the query with full attention
3. Select — top-scoring chunks after reranking become the final context

Configure the reranker in per-conversation RAG settings or globally. See Embedding & Reranking for available models.

---

Parent-chunk expansion

Retrieved chunks rarely exist in isolation. QARK fetches ±1 adjacent chunks from the same document, providing surrounding context. If chunk #14 scores highest, the model also sees chunks #13 and #15.

---

Citations

Every claim from your documents includes an inline citation badge (e.g., [1], [2], [3]).

Field	Content
Source document	Original file name
Page number	For paginated formats (PDF, DOCX, PPTX)
Section	Heading or chapter title when available
Chunk ID	Internal reference to the exact chunk
Relevance score	Similarity/reranking score (0–1)

Hover a badge for full citation details. Click to scroll to the referenced chunk in the document panel.

---

PDF image extraction

For PDFs with diagrams, charts, or embedded images:

• Enable in per-conversation RAG settings under Image Extraction
• Configure which vision-capable model handles image description
• Extracted images are converted to text descriptions and indexed as additional chunks
• Useful for technical manuals, slide decks, and research papers with figures

---

Per-conversation RAG configuration

Each conversation maintains independent RAG settings:

Setting	Options
Embedding provider & model	Any configured provider with embedding models
RAG generation model	Model for HyDE/step-back query generation
Reranker provider & model	Voyage AI or Jina AI reranking models
Direct injection threshold	1%–100% of context window (default: 30%)
Search strategy	Semantic, HyDE, Step-back, Auto
Image extraction model	Any vision-capable model

Access from the Info Panel → Config tab or the RAG config section in the conversation sidebar.

---

End-to-end workflow

1. Drag 5 research papers onto the conversation.
2. Watch each file progress through parsing → chunking → embedding → indexing → ready.
3. Ask: “What are the main differences in methodology between papers 2 and 4?”
4. Receive a structured comparison with inline citation badges linking to specific pages and sections.
5. Hover any citation to verify. Click to jump to the exact chunk.