Document Q&A¶

Document Q&A provides local retrieval-augmented analysis over uploaded PDF, TXT, and DOCX files.

Workflow¶

load documents
    -> extract/OCR text
    -> chunk
    -> embed
    -> retrieve
    -> apply grounding policy
    -> build Gemma request
    -> create_chat_completion()

🧲 Retrieval Controls¶

Chunks to Retrieve
Chunk Size
Chunk Overlap
Show Retrieved Chunks

These values govern retrieval, not grounding policy or response presentation.

🛡️ Grounding Controls¶

Require Grounding
Answer From Excerpts Only
Insufficient-Evidence Behavior

Supported insufficient-evidence behaviors include:

State Insufficient Information
Return Retrieved Excerpts
Best Supported Answer

🧮 Retrieval Backend¶

Bro can use:

Automatic
sqlite-vec
Cosine Similarity

Additional controls govern cosine fallback and whether the index is rebuilt for a query.

🗂️ Document Actions¶

The bounded action selector includes:

Answer Question
Summarize Active Document
Extract Key Points
Generate Outline
Extract Entities
Extract Tables
Compare Active Documents
Classify Document
Find Evidence
Generate Executive Summary
Extract Dates
Extract Organizations
Extract Requirements
Extract Action Items
Identify Contradictions
Identify Missing Information

📄 Document Parsing¶

Enable OCR
Prefer Native PDF Text
Include Page Markers
OCR Page Limit

OCR Page Limit is bounded rather than manually entered.

Vision-assisted OCR¶

When native PDF text is unavailable and OCR is enabled, Bro can render eligible PDF pages as images and route them through the Gemma multimodal runtime.

The OCR path therefore requires the same compatible mmproj configuration as Image to Text.

🔎 Diagnostics¶

Diagnostics are separated from parsing controls:

Show Diagnostics
Show OCR Status
Show Runtime Metadata

Response, inference, context, and runtime controls¶

Document Q&A uses the same conceptual ownership as Text Generation:

Response Controls describe output;
Inference Settings feed generation;
Context Controls govern context assembly;
Runtime Settings govern llama.cpp initialization.

This keeps the same session-state parameter from having different semantic homes in different modes.

Retrieved chunks¶

When enabled, Bro renders retrieved chunks with source information and score/distance metadata so the user can inspect the evidence base used by the answer.