Zero-Hallucination Legal & Financial RAG Pipelines: Line-Level Citation Grounding, Deterministic Entity Extraction, and Fast CPU Inference

Generic conversational LLMs are dangerous in a courtroom or financial audit room. If an advocate asks an ungrounded AI assistant to summarize a 60-page Petition or Commercial Plaint and the model invents a case citation, swaps the Petitioner and Respondent, or fabricates a prayer that does not exist on the record, relying on that output in court is catastrophic. In high-stakes legal and financial workflows, every single factual proposition must be anchored to a verifiable page and line coordinate that the practitioner can hover over to read verbatim and click to jump directly to the underlying document canvas.

When building KortiAI — the legal research and interactive workspace assistant inside KortiBundle (kortibundle.com) — we designed a hybrid Deterministic Structural Parser + Coordinate-Grounded RAG Pipeline that guarantees zero hallucinated parties or prayers, attaches clickable [Pg X · LY] badges to every finding, previews verbatim line quotes on hover, and scrolls the document canvas to the exact vertical baseline of the cited line when clicked.

1. Why Naive Chunk-Based RAG Fails on Legal Pleadings

Standard RAG tutorials split PDF text into arbitrary 500-token sliding windows and embed them into a vector database. On structured legal pleadings, court proceedings, and financial agreements, naive chunking fails in four predictable ways:

• Caption & Party Fragmentation: A Kenyan High Court or Magistrates' Court caption spans lines 1 to 16 of Page 1, with dot leaders (....................), multi-line party descriptions, and VERSUS separators. Naive chunking loses which party is above vs. below VERSUS, or confuses front-matter cover columns with the actual pleading caption.
• Loss of Line-Coordinate Provenance: Once paragraphs are concatenated into a 500-token string without per-line coordinate tags, the LLM can only cite "Page 2" instead of "Page 2, Line 15".
• Page Index vs. Effective Page Number Drift: In a compiled court bundle, Sheet #1 is the Cover Page, Sheet #2 is the Statutory Index, and Sheet #3 is Page 1 of the Petition. If the RAG pipeline confuses 0-based PDF sheet index with effective_page_number, hovering over [Pg 2 · L15] pulls Line 1 of the wrong sheet (often the generic header "REPUBLIC OF KENYA").
• Slow Tool-Call Routing on CPU Servers: Passing 2,000+ tokens of nested OpenAI JSON Schema definitions to a local CPU model causes 15-second prefill stalls.

2. Coordinate-Grounded Extraction & Hover Quote Verification

Instead of feeding raw unindexed text to the model, our pipeline indexes every horizontal text line on every non-front-matter pleading page with its exact effective_page_number, 1-based line_index, and vertical baseline_y coordinate in PDF points:

# Coordinate-Grounded Line Indexing & Multi-Line Prayer Extraction
from typing import List, Dict, Any

def build_coordinate_grounded_passages(page: Dict[str, Any]) -> List[Dict[str, Any]]:
    eff_pg = page.get("effective_page_number")
    sheet_idx = page.get("page_index", 0)
    pg_label = f"Page {eff_pg}" if eff_pg is not None else f"Sheet {sheet_idx + 1}"

    grounded_hits = []
    for dl in page.get("detected_lines", []):
        snippet = (dl.get("text_snippet") or "").strip()
        if not snippet:
            continue
        line_no = dl.get("line_index", 1)
        grounded_hits.append({
            "sheet_index": sheet_idx,
            "effective_page": eff_pg,
            "page_label": pg_label,
            "line_index": line_no,
            "baseline_y": dl.get("baseline_y", 0.0),
            "cite_tag": f"[{pg_label}, Line {line_no}]",
            "snippet": snippet,
        })
    return grounded_hits

When the backend streams a response to the frontend chat drawer, it also emits a verified sources array containing the exact verbatim text of every cited (effective_page, line_index) pair. When an advocate hovers over [Pg 2 · L15] in KortiBundle, the UI immediately renders a floating quote card showing the exact text of Page 2, Line 15—and when clicked, calculates yRatio = baseline_y / page.height to scroll the viewport directly to that line and illuminate it with a golden highlight bar.

3. Fast JSON Intent Routing Without Brittle Regex Gates

When a user types an ambiguous workspace command like "toggle view into edit mode", "move page 2 one step up", or "get rid of page 4", relying on rigid regular expressions breaks whenever the user adds a natural phrase like "one step" or uses a synonym like "toggle". Conversely, passing a 2,100-token OpenAI tool schema to a local CPU model causes unacceptable latency. By using a compact 280-token format: "json" system prompt kept permanently warm in the LLM's KV prefix cache, intent classification completes in 1.2 to 2.0 seconds on CPU with near-zero classification errors.