Raw documents degrade AI retrieval quality.

Green Energy and Sustainable Finance organizations hold critical data in raw documents like PDFs that were only designed for human consumption, not machine retrieval.

IRENA’s Renewable Power Generation Costs Report is a perfect example of the type of pre-parsed document that would be provided by a client for use in the RAG layer of an AI system. It contains mixed text and data tables, a multi-section layout, technical terminology, footnotes, and statistical figures.

The problem is when this type of raw document is fed directly into a RAG system, it produces artifacts that pollute an AI agent’s vector database causing it to hallucinate.

Precision audit conducted before a single vector is written.

Funnel Amp applied a structured RAG ingestion audit to the source document. Using Docling for intelligent document parsing to preserve tables, headings, and reading order—every section was manually classified against a purpose-built two-tier metadata schema.

A domain-specific Controlled Vocabulary Tag (CVT) set was derived directly from the source document, ensuring retrieval precision for Green Energy analysts. Noise chunks were flagged and excluded before any embedding occurred.

The result is a retrieval-ready corpus where every chunk carries full provenance, domain context, and technology classification.

Clean documents alone do not produce a working AI system.

Most RAG architecture applies a single default chunking strategy and a generic embedding model without any analysis of whether those choices are appropriate for the corpus. The result? Silent degradation. Critical content gets truncated before it ever reaches the vector database, retrieval scores suffer, and AI-generated answers lose accuracy without any visible error to investigate.

For Sustainable Finance clients working with dense regulatory frameworks like the ICMA Green Bond Principles, this is particularly consequential. A compliance officer receiving a degraded answer about reporting requirements or eligible project categories has no way of knowing the underlying system failed at the architecture level.

Every architectural decision tested, measured, and documented.

A complete RAG pipeline was built using the ICMA Green Bond Principles 2021 as the source corpus. Instead of relying on a default approach, a careful comparison of two chunking strategies revealed that structure-based chunking would silently truncate critical regulatory content in 3 of 23 sections, including the cornerstone “Use of Proceeds” section.

Sentence-window chunking with deliberate overlap was selected instead. The embedding model selection was driven by corpus analysis, with all-mpnet-base-v2 being chosen after discovering that 278-token chunks would exceed the default model's 256-token ceiling. The completed pipeline was validated with live query testing against ChromaDB, achieving relevance scores above 0.80 on targeted Sustainable Finance queries.

Most RAG systems are deployed without any way to measure whether they work.

Organizations invest in building AI knowledge systems but have no formal methodology for evaluating whether those systems retrieve the right information accurately. Without a structured evaluation framework, retrieval failures and hallucinations go undetected until they surface as errors in front of end users relying on the system for real decisions.

There is no error message when a RAG system retrieves the wrong chunk. There is no warning when a generated answer includes information the retrieved content does not support. The absence of measurement means problems are discovered by users, not architects, and by then, trust in the system is already eroding.

Systematic evaluation across two domains with documented evidence at every step.

Funnel Amp designed and executed a complete RAG evaluation framework across two corpora: the IRENA Renewable Power Generation Costs 2023 report (Green Energy) and the ICMA Green Bond Principles 2021 (Sustainable Finance). A ten-question golden evaluation dataset was purpose-built from primary source documents, with each question paired with a known correct answer, key verification phrases, and a documented retrieval trap risk assessment.

Two retrieval configurations were tested against the same dataset. Hit Rate and MRR (Mean Reciprocal Rank) were calculated per domain and overall. A retrieval coverage failure was identified and diagnosed as a chunking architecture gap rather than a question design weakness. Metadata-assisted retrieval was implemented to resolve semantic collision between adjacent sections, reducing the retrieval trap performance gap from 33% to 8%.