Search & Chat

The preprocessor includes two modes for finding and understanding your documents after they have been processed and saved.

Search

Find documents by keyword with type and date filters. Results show thumbnails, parsed filename fields, and highlighted OCR text.

Chat

AI assistant powered by Gemini. Reads documents, queries databases, checks Monday.com, runs calculations, and analyzes uploaded files.

Both modes are accessible from the navigation sidebar.

---

Search

Search uses a hybrid retrieval system that combines vector similarity (semantic meaning) with BM25 keyword matching. Results are ranked by a fused relevance score.

Running a Search

1. Type your query into the search bar.

2. Filter by type using the dropdown to the right of the search bar: All, Invoice, or Document.

3. Press Search or hit Enter to run the query.

4. (Optional) Enable the date filter by toggling the switch below the search bar. Set the From/To range to restrict results to documents whose filename date falls within that window. Documents without a parseable date are still included.

Search Results

Each result card contains:

Element	Description
Thumbnail	Preview of the matched page. Click to open a full page preview dialog with page navigation.
Filename fields	Parsed from the filename — property, vendor, and invoice number for invoices; doc type and property for documents.
Type & page	Document type badge and which page matched the query.
Location	Server file path where the document is stored.
Matched text	Snippet of OCR text with search terms highlighted in yellow.
Score	Relevance score, color-coded: green (70%+), orange (40—69%), red (below 40%).
Open File	Launches the PDF from the server in your default viewer.

Click Load More Results at the bottom to fetch additional matches.

Search Sidebar

The sidebar shows a green Server connection active label when the server is reachable (or a red warning if not), followed by index statistics: total indexed pages, total documents, and a breakdown by type. If the old server data source is active, its page count appears separately.

Button	Action
Sync Search Data	Pulls server records into your local index and pushes local-only records back. Only visible when connected.
Rebuild BM25	Rebuilds the keyword search index from ChromaDB data. Run after syncing or if keyword results seem stale.
Export Data	Exports the full index to a JSON file for backup or inspection.
Clean Up DB	Scans for entries pointing to missing or local-only files and offers to remove them. Only visible when connected.

---

Chat

Document Chat is an AI assistant that plans multi-step workflows automatically based on your question.

How It Works

1. Intent classification — Gemini analyzes your question and plans which data sources to query and in what order.

2. Retrieval — For document queries, it expands your question into multiple search queries, runs hybrid search, and fuses results using Reciprocal Rank Fusion (RRF).

3. Reranking — Large result sets are filtered by an LLM reranking pass to remove noise.

4. Answer generation — Gemini reads the retrieved content and generates a response with full conversation history for context.

Every response shows a collapsible processing steps status bar. While running it displays each step live; when finished it collapses to a checkmark with “Complete.” Expand it to see which sources were queried, how many documents were found, and the full workflow plan.

Data Sources

The chat pulls from five sources and combines them automatically when a question spans multiple.

Documents

Searches the vector index of OCR’d PDF pages. Use for anything written in a document: amounts, dates, terms, account numbers, lease clauses, notice content.

Database

Queries SQLite tables (entity, vendor, mortgage, invoice_tracker). Use for structured lookups: vendor lists, entity/property info, owner/manager details, utility flags.

Monday.com

Reads the AP board for invoice workflow: status, due dates, paid dates, recurring flags, group assignments. Also reads the M&O board for task notices.

Upload

Reads a file you attach to the chat. Use for reconciliation, comparison, or analysis of external data against your system.

Code

Runs Python in a read-only sandbox with pandas, SQLite access, and data from prior pipeline steps. Use for calculations, aggregations, charts, and visualizations.

Chat Input

The chat input bar sits at the bottom of the page. On the left is a paperclip icon for attaching files, and on the right is a send arrow. Type your question and press Enter or click the send arrow.

Uploading Files

Click the paperclip to attach a file before or while typing your message. Accepted types: PDF, CSV, Excel (.xlsx/.xls), and TSV.

Note

Uploaded files stay active for the full chat session. The sidebar shows the attached file under ATTACHED FILE. Attaching a new file replaces the previous one. Clearing the chat also clears the attachment.

The system knows to use the uploaded file when you reference “the file,” “the spreadsheet,” “the register,” or similar language. You can also just attach a file without typing anything and it will analyze the contents automatically.

Example Queries

Lookups

Single-source queries that search documents, the database, or Monday.com directly.

• “What is on the latest SCE invoice for Westwood?”
• “Find the lease for 1234 Main Street”
• “Show me all task notices from this month”
• “What is the rent amount on the LAX lease?”
• “List all utility vendors”
• “Which properties does owner EM manage?”
• “What invoices are pending on the AP board?”
• “Show me all paid invoices from last week”

Multi-Step & Upload

The system automatically chains sources when your question requires it. For upload queries, attach a file first using the paperclip.

• “Find the most expensive invoice from vendor code AAPL” — resolves vendor name from the database, then searches documents
• “Which utility vendors have invoices on the Monday board?” — pulls utility list from database, then checks Monday
• “Show me leases for entities managed by owner EM” — queries the entity table, then searches documents for those properties
• “Find the latest Monday invoice and tell me the amount” — gets the item from Monday, then finds the PDF and extracts the dollar amount
• “Reconcile this check register with the Monday AP board” — reads your uploaded CSV, then pulls Monday data to compare
• “Match the uploaded vendor list against our vendor database” — parses your file, then queries the vendor table
• “What vendors in this file are missing from our system?” — compares uploaded data against the database

Code & Charts

Analytical queries that run Python in the sandbox. When financial data lives in documents rather than the database, the system extracts it from PDFs first then passes it to code.

• “What percentage of vendors are utilities, broken down by owner?”
• “Show a month-by-month invoice count trend for 2025”
• “Chart the average invoice amount per property”
• “Which properties have the most vendors and what are the top 3 per property?”
• “Show a visualization of LAX rent prices over time” — extracts data from lease documents first, then charts it
• “Compare rent amounts across Westwood properties” — finds leases, then builds a comparison visualization

Bulk CSV Extraction

When you ask for data across many documents in a tabular format, the system automatically switches to bulk mode. It casts a wider search net, processes documents in batches, and produces a downloadable CSV.

For example: “Give me a CSV of all vendor invoice amounts for 2024” or “Extract account numbers from all SCE invoices.”

Response Features

Formatted Tables

Responses often include markdown tables for structured data like property overviews, rent schedules, vendor lists, and invoice summaries.

Referenced Files

Expandable list of documents the AI used. Shows filename, type, matched page, relevance score, and server path. Click Open to launch the file.

Download CSV

Appears when the response includes tabular data. Downloads the extracted data as a CSV file with a generated filename.

Export Response

Downloads the individual question and response as a formatted PDF.

Retry

Appears on the most recent response. Reruns the full pipeline for the same question, which can produce different results.

Charts

When the code execution step generates a visualization, it renders inline below the text response.

Chat Sidebar

The sidebar shows system status at a glance.

Section	What it shows
Search Index	Pages, documents, invoice count, and document count from the vector database.
Database	Entity, vendor, and tracked invoice counts from SQLite.
Connections	Status indicators (checkmark or X) for Server, Monday.com, and ChromaDB.
Deep Thinking	Checkbox that increases reasoning depth for all Gemini calls. Useful for complex questions; responses take longer.
Document Read Depth	Controls how much OCR text per page the AI reads. Four presets: Quick (minimal text, fast), Standard (balanced, default), Thorough (more text, better for detailed extraction), and Full (entire document, no truncation). Higher settings improve accuracy on long or dense documents but use more tokens and may require additional processing batches.
Attached File	Name and size of the currently uploaded file, if any.
Clear Chat	Resets the conversation, clears the attached file, and resets token tracking.
Export Chat	Downloads the full conversation as a formatted PDF. Appears after at least one exchange.

Tip

For the best results, be specific. “Show me the lease for 1234 Main St” works better than “find a lease.” You can also name the source directly: “check Monday for pending invoices” or “search the database for utility vendors.”

When to change read depth

Use Quick for simple lookups where you just need a name, date, or status. Use Thorough or Full when extracting detailed data from long documents like leases or multi-page invoices. The processing steps status bar will note the active read depth when it differs from Standard.

Requirements

Chat requires a Google API key configured in App Settings.