CodeBailey products, built to move fast and hold up.
A purpose-built product for every step of the flow. Pick only what you need and combine them.
How it flows
Four stages, five product groups.
Seventeen Bailey products are organized into collection, parsing, knowledge, agents and operations, so you can pick only what you need and connect it to existing systems.
Change-data sync from databases and ERP, scheduled crawling of disclosures and bid notices, and mail grouped by task, all with a source-tracked collection history.
CDC sync
Bid & disclosure crawling
Source history
Bailey ConnectConnects databases, ERP, groupware and SaaS, syncing only what changed in real time.Connectors attach to your databases, ERP, groupware and SaaS to gather data, and from then on change data capture (CDC) syncs only what changed. Only the changed rows of ERP order and inventory tables reach your analytics database, so nothing is copied over in full each time. Every record keeps its source and history, so you can always trace where a value came from.CDCKafkaAirflow
Key capabilities
Connectors for major databases, ERP and groupware
Incremental sync via change data capture (CDC)
Collection history and source tracking
Use case
Stream only the changed rows of ERP order and inventory tables into your analytics database.
What you get
A unified store with source-tracked collection history
Bailey CrawlerAn agent that collects web pages, portals, disclosures and boards on schedule.Collects websites, portals, disclosures and boards on the schedule and rules you set. It compares against what it already has, filters duplicates and flags only what changed, and on a failed crawl it sends an alert and retries. It suits work that has to check disclosures and bid notices at the same time every day.SchedulerDiffRetry
Key capabilities
Schedule- and rule-based collection
Duplicate and change detection
Failure alerts and retries
Use case
Collect disclosures, bid notices and boards on schedule and flag only what changed.
What you get
Change-detection results and a crawl log
Bailey MailGathers mail, messenger threads and attachments and groups them by task.Gathers mail, messenger threads and attachments and groups them automatically by task and counterpart. Contract and invoice mail, for example, is organized by counterpart, and attachments go straight to parsing so their contents can be read too. It cuts the work of sorting scattered inboxes by hand.MailAttachmentsGrouping
Key capabilities
Mail and attachment collection
Automatic grouping by task and counterpart
Hand-off to parsing
Use case
Group contract and invoice mail with attachments by counterpart and pass them to parsing.
Structure-preserving parsing, OCR and field extraction.
Three parsing products.
Merged-cell HWP tables and scanned contracts are read with their structure intact, and amounts, deadlines and parties are passed on as validated data.
HWP tables
OCR
Field validation
Bailey ParserConverts PDF, HWP and Office files while keeping headings, tables and layout.Reads HWP, PDF, DOCX and XLSX files with their heading hierarchy, tables and footnotes intact instead of extracting plain text. Merged-cell tables in HWP reports keep their cell positions, so table contents don't break in search and analysis. The output is structure-preserving JSON that the embedding and agent stages pick up directly.HWPPDFDOCXXLSX
Key capabilities
HWP, PDF, DOCX and XLSX support
Table rows, columns and merged cells preserved
Heading hierarchy and footnotes kept
Use case
Turn merged-cell tables in HWP reports into data with rows and columns intact.
What you get
Structure-preserving JSON
Bailey OCRRecognizes scanned documents, images, drawings and handwriting.Recognizes text in scanned PDFs and images, and for documents divided into regions, such as drawings and forms, it finds the region positions as well. Handwritten notes are recognized too. The output is text with region positions that the parsing and extraction stages use next.OCRLayoutHandwriting
Key capabilities
Text recognition for scanned PDFs and images
Drawing and form region detection
Handwriting recognition
Use case
Recognize text and regions in scanned contracts, drawings and handwritten notes.
What you get
Text with region positions
Bailey ExtractExtracts and validates fields such as contract amount, deadline and parties, then classifies and summarizes.Pulls predefined fields such as contract amount, deadline and parties out of documents and validates formats like dates and amounts. It also classifies the document type and passes along a summary of the key points. It suits work that checks the same fields again and again, such as contract review.SchemaValidationSummary
Key capabilities
Field extraction with format validation
Document type classification
Key-point summaries
Use case
Pull amount, due date and parties from contracts, validate formats and classify the type.
Products that turn documents into sourced knowledge.
Embed, retrieve and answer with citations.
Three knowledge products.
Rulebooks are indexed clause by clause, hybrid search finds the evidence, and answers cite their sources using only documents the user is permitted to see.
bge-m3
pgvector HNSW
Permission filter
Bailey EmbedStructure-aware chunking and multilingual embedding with bge-m3.Splits documents along their structure, such as headings and clauses, rather than at a fixed length, then turns each chunk into a 1024-dimensional vector with the multilingual bge-m3 model. Frequently revised documents like rulebooks re-index only the changed clauses instead of rebuilding everything. These chunks and vectors are the basis for search and source citations.bge-m31024-dChunking
Key capabilities
Structure-aware chunking
Multilingual embedding with bge-m3
Partial re-indexing when documents change
Use case
Split a rulebook by clause and re-index only the revised clauses.
What you get
Chunks and vectors
Bailey VectorRetrieval with pgvector HNSW indexes and hybrid search.Finds vectors quickly with pgvector HNSW indexes and fuses the results with keyword (BM25) search using RRF, so fewer documents are missed than with either method alone. The fused candidates are reranked by a cross-encoder and handed on in order as evidence for the answer.pgvectorHNSWBM25RRF
Key capabilities
pgvector HNSW vector index
BM25 + vector hybrid search (RRF)
Cross-encoder reranking
Use case
Combine keyword and semantic search to miss fewer documents, then rerank.
What you get
A ranked list of evidence
Bailey RAGAn air-gapped RAG platform that answers from your documents and cites its sources.Runs inside an air-gapped network on Ollama, answers from your documents and attaches the source document and passage to every answer. Access filters by document classification keep documents a user may not see out of the results, and every query and upload is written to an audit log.OllamaCitationsAudit log
Key capabilities
Source document and passage cited for every answer
Access filters by document classification
Audit log of queries and uploads
Use case
Answer policy questions with the clause cited, excluding documents outside the user's permissions.
Design agent roles on screen, connect internal systems as MCP tools, and start right away with ready-made agents for reports, complaints and policy review.
MCP tools
Human approval
Task templates
Bailey Agent StudioDesign agent roles, tools and flows on screen.Define each agent's role and permissions, the tools it may use and how work passes between agents, all on screen. For example, complaint handling can be designed as intake, research, drafting and review agents. Every change to a design is kept as a version, so you can track what changed and when.RolesFlowVersioning
Key capabilities
Role and permission definitions
Tool wiring and flow design
Versioned designs
Use case
Design complaint handling as intake, research, draft and review agents on screen.
What you get
Versioned agent designs
Bailey OrchestratorSplits work across agents and collects the results.Splits a large job across several agents and gathers the results into one. If an agent stops, its task state is handed to another agent to carry on, and key steps wait for human approval before going ahead. Completed tasks keep an approval record.RoutingTask queueHuman approval
Key capabilities
Task distribution and routing
Work redistribution and state hand-over on failure
Human approval steps
Use case
Distribute work across agents and, if one stops, hand its state to another to carry on.
What you get
Completed tasks with an approval record
Bailey ToolsConnects internal systems and APIs as MCP tools.Exposes internal functions such as ERP lookups, approval requests and mail sending as MCP servers that agents can run as tools. Each tool has its own permissions and every call is logged, so you can see which agent did what and when.MCPPermissionsCall log
Key capabilities
Internal APIs exposed as MCP servers
Per-tool permissions and call logs
Lookup, create and send tools
Use case
Expose ERP lookups, approval requests and mail sending as MCP tools agents can run.
What you get
A permissioned set of internal tools
Bailey TemplatesReady-to-use agents for reporting, civil-complaint handling and policy review.Provides common work such as report writing, complaint handling and policy review as ready-to-use agents. You can start just by connecting your documents, and adjust the template when your process differs. Not having to design from scratch makes adoption faster.ReportComplaintPolicy review
Key capabilities
Flow templates per task
Start by connecting your documents
Adjustable to your process
Use case
Start report writing, complaint handling or policy review just by connecting your documents.
Domain fine-tuning, before-and-after evaluation, LLM serving on separated networks, and attack blocking with PII masking cover the full operating cycle.
LoRA & DPO
RAGAS evaluation
Bailey Guard
Bailey TuneLoRA and DPO fine-tuning on your domain data.Builds training and evaluation datasets from work documents and query logs, then adapts the model to your domain with LoRA or QLoRA, which train only a small part of the parameters and fit on a single GPU. DPO aligns answers with the style you prefer, and the result is kept as a domain adapter and a training record.LoRAQLoRADPO
Key capabilities
Training and evaluation dataset building
LoRA / QLoRA fine-tuning
DPO preference alignment
Use case
Build datasets from work documents and query logs and adapt a model on a single GPU.
What you get
A domain adapter and training record
Bailey EvalMeasures accuracy and faithfulness with RAGAS and compares before and after changes.Measures the faithfulness and accuracy of answers with RAGAS. A managed evaluation set lets you compare before and after any model or prompt change on the same questions, and regression tests catch where quality dropped. Results are kept as a before-and-after report.RAGASEval setRegression
Key capabilities
RAGAS faithfulness and accuracy metrics
Evaluation set management
Before/after regression tests
Use case
Compare before and after any model or prompt change on the same evaluation set.
What you get
A before-and-after report
Bailey ServeServes LLMs in air-gapped and on-premise environments.Serves LLMs with Ollama in air-gapped and on-premise environments and manages GPU resources. Models are versioned, so you can swap a model without changing the apps that use it. Internal apps connect through an in-house LLM API.OllamaGPUVersioning
Key capabilities
On-premise serving with Ollama
GPU resource management
Model swaps and versioning
Use case
Serve LLMs on a separated network and swap models without changing apps.
What you get
An internal LLM API
Bailey GuardBlocks prompt attacks, masks personal data and logs every use.Stops attacks such as prompt injection at the input and automatically masks personal data like resident registration numbers and phone numbers. Every use is written to an audit log and usage is monitored, so you can check later who used it and how.GuardrailPII maskingAudit log
Key capabilities
Prompt-injection defense
Automatic personal-data masking
Audit log and usage monitoring
Use case
Block prompt injection, mask resident numbers and phone numbers, and log every use.
