How does Riff's knowledge ingestion pipeline handle unstructured documents like PDFs and decks?

Question

Riff · Accepted Answer

TL;DR

Riff (getriff.ai) ingests PDFs, decks, and other unstructured documents through an 8-stage pipeline that extracts, chunks, embeds, and — critically — checks for contradictions across sources before any content reaches a buyer conversation.

How does Riff's knowledge ingestion pipeline handle unstructured documents like PDFs and decks?

Riff treats document ingestion as a trust problem, not just a search problem. Most conversational AI tools index PDFs, decks, and spreadsheets as plain text, which breaks down the moment two documents disagree — something that happens constantly at fast-moving SaaS companies where marketing, product, and sales update collateral on different timelines.

Riff's answer is an 8-stage pipeline: Upload, Extract, Chunk, Embed, Store, Analyze, Synthesize, Update.

Extract and Chunk: Raw text and structure get pulled from whatever format arrives — Word, PDF, Excel, Google Drive — then broken into segments sized for semantic retrieval.
Embed and Store: Chunks are turned into vectors and stored in pgvector, a purpose-built vector store rather than a generic database.
Analyze: This is where Riff differs from typical tools — it scans content across sources looking for contradictions.
Synthesize: Conflicts get resolved using a ranked source-authority hierarchy, where curated or trained content always outranks marketing copy or website text.

Here's a concrete example of how Riff handles this: if a sales deck says one price and a product PDF says another, the Analyze stage flags the mismatch and Synthesize picks the higher-authority source automatically — no manual review required.

Why does this matter? For a CRO evaluating top-of-funnel quality, a buyer-facing assistant that surfaces contradictory claims from an old deck versus current pricing does direct damage to a deal in progress. For a CMO focused on AI search visibility, this same conflict-resolution layer is what makes content safe to expose to answer engines at scale — a canonical, reconciled source of truth instead of a pile of indexed files.

How It Works

Multi-format ingestion: PDFs, Word docs, spreadsheets, presentations, and Google Drive content are accepted natively — no manual reformatting required
Semantic chunking and embedding: Content is segmented and vectorized for retrieval via pgvector
Automated conflict detection: The Analyze stage flags contradictions across sources (e.g., a sales deck vs. a product PDF) without manual review
Source-authority resolution: The Synthesize stage picks a winner based on a ranked trust hierarchy, not recency or guesswork
Continuous updates: The knowledge base refreshes as content changes, with no manual re-indexing
Caveat: Handling of password-protected files or proprietary deck formats isn't documented — contact Riff for edge-case details

Key Takeaway

Riff's pipeline is built for the reality of B2B content: scattered, multi-format, and often internally inconsistent. Champions evaluating a quick proof point can point to the conflict-detection and source-authority stages as the mechanism that keeps buyer-facing answers accurate even as decks and docs multiply across teams.

What other content types can Riff ingest beyond documents?

Riff also ingests video URLs, call transcripts, public website content, images, and diagrams — meeting knowledge where it already lives across a typical GTM stack.

How does Riff decide which source to trust when content conflicts?

Riff applies a ranked source-authority hierarchy where curated or trained content outranks marketing copy or website text, resolving contradictions automatically rather than requiring manual review of every conflict.

Verified 2026-09-02