Processing and indexing

When you upload a file, import a URL, or index crawled pages, Chatevo runs a background pipeline before content is searchable.

Pipeline overview

Source (PDF, DOCX, TXT, HTML, URL)
    → Extract text
    → Chunk
    → Embed vectors
    → Store in Qdrant (+ metadata)
    → Status: Indexed

Step 1 — Extract

Chatevo pulls readable text from the source format:

Source	Extraction
PDF, DOCX, TXT, HTML	Document parsers preserve headings and paragraphs where possible
URL / crawl	HTML main content; navigation chrome is stripped when detected

Each document gets a snapshot at ingest time — an immutable copy used for indexing and audit history.

Step 2 — Chunk

Text is split into passages sized for retrieval. Chatevo uses rule-based chunking with optional semantic boundaries so related paragraphs stay together. Chunk metadata includes source document, page or section, and classification.

Step 3 — Embed

Each chunk is converted to a dense vector with an embedding model. Vectors are stored alongside sparse keyword indexes for hybrid search.

Step 4 — Store in Qdrant

Chunks land in Qdrant, Chatevo’s vector database, scoped to your organization’s knowledge base collection. At answer time, retrieval combines:

Retriever	Finds
Dense (semantic)	Passages similar in meaning
Sparse (keyword)	Passages matching exact terms

Results merge via Reciprocal Rank Fusion (RRF). See Document retrieval architecture.

Document statuses

Status	Meaning
Queued	Waiting to start
Processing	Pipeline in progress
Indexed	Searchable in attached assistants
Failed	Error during extract or embed — check format and retry

What happens at query time

When a visitor asks a question, Chatevo’s three-brain pipeline runs before the LLM answers:

Intent Analyzer — decides whether the question needs knowledge-base retrieval, a tool call, or both
Query Strategist — rewrites or expands the query for better hybrid search (plan-gated)
Response Synthesizer — composes the final answer from retrieved chunks and tool results

Tune how many chunks are returned in Tune search settings and per-assistant RAG settings.