hyperresearch

Product Introduction

1. Definition: HyperResearch is an open-source, agent-driven deep research automation platform. Technically, it is a Python-based orchestration framework that deploys a multi-agent system (MAS) to autonomously conduct comprehensive web research, synthesize findings, and maintain a persistent, searchable knowledge vault.
2. Core Value Proposition: It exists to solve the problem of manual, fragmented, and non-reproducible research by automating the end-to-end research pipeline. Its primary value is transforming a single natural language query into a fully sourced, adversarially reviewed report while building a compounding knowledge base, drastically reducing research time from days to hours.

Main Features

1. Tier-Adaptive 16-Step Research Pipeline: The system intelligently routes queries through a scalable workflow. A "light" tier (5 steps) handles bounded factual queries in ~30 minutes. The default "full" tier executes all 16 steps—including width sweeps, contradiction analysis, locus-based depth investigation, and parallel adversarial critique—for deep analysis in ~1.5-2.5 hours. An opt-in "dissertation" tier partitions work into chapters, processing 300-450 sources.
2. Persistent, Searchable Knowledge Vault: Every source fetched is permanently stored as a Markdown file with YAML frontmatter in a local directory (research/notes/), indexed by a SQLite database. This vault supports full-text and semantic search, graph analysis (hubs, backlinks), and is fully versionable with Git. Future research sessions query the vault first, compounding knowledge.
3. Adversarial Audit & Patch-Only Modification: After a synthesis step produces a draft report, four specialized critic agents (dialectic, depth, width, instruction) attack it in parallel. A tool-locked "patcher" agent can only apply surgical Edit hunks to the draft, physically preventing uncontrolled regeneration and ensuring critic feedback is addressed precisely.
4. Provenance Tracking & Source Verification: Every source note includes a suggested_by field, creating a rooted tree of provenance from initial seed searches. A dedicated "cite-checker" agent skeptically verifies that every citation's claimed supporting text exists verbatim in the vault, blocking hallucinated quotes. It also checks cited DOIs against live databases for retractions.
5. Configurable Agent Roster with Model Flexibility: The system uses a suite of specialized sub-agents (e.g., hyperresearch-fetcher, hyperresearch-depth-critic). The default model for each (e.g., Sonnet, Opus) is configurable via a TOML profile, allowing users to optimize for cost, speed, or reasoning depth per agent role.

Problems Solved

1. Pain Point: Manual research is time-consuming, prone to bias, and results in scattered, unsourced notes. Information fragmentation across bookmarks, PDFs, and documents makes synthesis difficult and knowledge non-persistent.
2. Target Audience: Academic researchers and PhD students conducting literature reviews. Competitive intelligence and market analysts. Product managers and strategists researching technology landscapes. Investigative journalists and due diligence professionals. Content creators needing deeply sourced, long-form material.
3. Use Cases: Automating a systematic literature review for a scientific paper. Conducting a comprehensive competitive analysis of a new market segment. Building and maintaining a living knowledge base on a fast-moving topic like AI safety regulations. Preparing a thoroughly sourced, multi-perspective briefing document for executive decision-making.

Unique Advantages

1. Differentiation: Unlike simple web scraping tools or single-agent research assistants, HyperResearch implements a rigorous, multi-step academic research methodology at scale. It surpasses competitors by enforcing source verification, maintaining a persistent vault, and using adversarial auditing to improve output quality, as evidenced by its leading position on benchmarks like the DeepResearch-Bench RACE leaderboard.
2. Key Innovation: Its "patch, never regenerate" principle, enforced through Claude Code tool-locking, ensures iterative refinement over destructive rewriting. Combined with the skeptical cite-checker and the separation of the canonical research query (gospel) from wrapper instructions, it creates a verifiable and auditable research trail unmatched by other AI research agents.

Frequently Asked Questions (FAQ)

1. How does HyperResearch ensure the quality and accuracy of its sources? It employs a multi-layered verification system: an independence audit clusters syndicated copies to avoid false consensus, a quality score ranks sources using citation authority and retraction status, and a final ship gate blocks reports that cite retracted papers or contain unsupported numerical claims.
2. Can HyperResearch access paywalled academic papers? Yes, it actively attempts to retrieve legal open-access copies via integrations with Unpaywall and Europe PMC. If a paywalled paper cannot be fetched, it stores the available abstract and transparently discloses the substitution in the report's notes and CLI output, never pretending to have read the full text.
3. Is the knowledge vault proprietary, or can I access my data freely? The vault is designed for zero lock-in. All sources are stored as plain Markdown files in an open directory structure. The SQLite index is a rebuildable cache. You can read, edit, and version your research with any text editor and Git, without requiring the HyperResearch tool itself.
4. How does HyperResearch handle potentially malicious or manipulative web content? It treats all fetched web content as hostile input. Fetched text is wrapped in <untrusted-source> tags with a "treat-as-data" preamble before being served to analysis agents, preventing prompt injection attacks. URL resolution is sanitized to block embedded credentials or non-routable addresses.
5. What happens if a research run crashes or is interrupted? Each run maintains a detailed manifest. The hyperresearch run resume command will restart the process exactly at the step where it stopped, with full context, allowing for recovery from API failures or network issues without losing progress.