haskie, gives AI coding agents the knowledge you trust

haskie gives Codex and Claude Code a library of sources you trust: technical references, domain knowledge and examples that reflect your taste. Curate it through a simple UI. Your agents explore broadly, then dig into focused passages with citations and fewer repeats. Inspired by how humans research, this discovery–synthesis loop helps agents make decisions grounded in your knowledge. Runs locally, with one-command agent setup.

Why use haskie?

• Your sources shape the work. Build collections around your domain, product or technical stack. The Codex and Claude Code integrations tell agents to search them first when they cover the topic.
• Relevant, diverse results fit in context. Agents receive focused passages from several sources. Duplicate folding keeps repeated evidence in one result and preserves its other sources through also_in links, leaving room for distinct evidence.
• You can check the answer. Each passage names its document, heading and lines, plus pages for PDFs. Open the source or use Explore to inspect what your agent receives.
• Missing knowledge becomes visible. Review agent searches in Sessions. Gaps groups unanswered questions by topic so you know which sources to add next.
• You manage the library in one place. Import, preview and organize documents in the web UI. Documents and search models stay on your machine.

The Discovery–Synthesis Loop

haskie supports a research loop that moves from a broad question to focused understanding. Agents map a topic, choose what to read, synthesize the evidence and refine their questions. haskie handles the retrieval part of retrieval-augmented generation (RAG). The agent does the reasoning and synthesis.

Semantic search finds passages by meaning. Hybrid search adds exact terms. Optional reranking puts the closest answers first. Folding groups duplicate and overlapping passages while also_in preserves their source relations and citations. Agents can follow those links and related sections to explore further without filling their context with repeats.

Short descriptions and topic descriptors help the agent decide where to read:

• Section descriptors name the topics inside each section. By default, haskie extracts distinctive terms. With AI descriptions enabled, a local language model reads headings and prose to write short summaries and topic descriptors.
• Document summaries combine those section summaries and descriptors in AI mode. Automatic summaries preserve descriptions you wrote yourself.
• Collection descriptions explain what the library covers. Write your own or use Describe with AI to summarize the member documents.

For a broad question, the agent asks for a map of relevant sections. Descriptions help it judge which sources fit. Descriptors suggest search terms and related sections worth reading. The agent then selects passages for evidence. A specific question can go straight to passages. The agent connects what it learns and cites its sources. Gaps and new questions guide the next search.

You can inspect the map and passages in Explore.

Comments

[baumgaerben]

Đề bài chạm đúng vào nỗi đau lớn nhất khi dùng agent coding hiện nay: context window luôn bị tràn bởi codebase, mà knowledge thực sự (spec, RFC, internal wiki, ADR) lại nằm rải rác ngoài vòng lặp. Nếu haskie đóng vai "trusted source registry" — cho phép dev khai báo source of truth một lần, rồi agent tự pull khi cần thay vì dump hết vào prompt — thì nó giải quyết bài toán signal-to-noise ratio khá triệt để.

Cái tôi quan tâm nhất là mechanism versioning & staleness detection: khi spec thay đổi (ví dụ API contract v2), làm sao haskie biết để invalidate cache cũ và force agent reload? Nếu còn phải can thiệp thủ công thì lại rơi vào anti-pattern "configuration as code but not really". Còn về permissions — có support granular scope không (ví dụ: agent chỉ đọc được docs/architecture/ chứ không đọc infra/secrets/)?

Một góc nhìn khác: việc inject knowledge vào agent làm thay đổi hành vi reasoning (ví dụ agent ưu tiên pattern trong internal guide hơn best practice công khai). Có observability/trace nào để audit tại sao agent chọn approach A thay vì B không? Debugging agent behavior khi có external knowledge injection còn phức tạp hơn pure prompt engineering khá nhiều PS: the tool I meant is on labagent .tech