SaaS Metrics
Gemini Executive Synthesis
Performance degradation and excessive token usage in long-form content generation due to 'full context injection'
Technical Positioning
Optimizing LLM context management for scalability and efficiency in long-form content generation
SaaS Insight & Market Implications
`inkos` experiences severe performance degradation in long-form writing, with single-chapter generation times reaching 40 minutes. This is attributed to 'full context injection' where `spot-fix`, `Reviser`, and `Settler` phases feed entire project contexts, including the `chapter_summaries.md` file, to the LLM. This file grows excessively, leading to high token usage, slow responses, and model 'thinking failures.' The pain point is the lack of intelligent context pruning, making the system economically unviable and functionally impractical for extended projects. Market implication: Scalability in AI content generation hinges on efficient context management. Solutions must move beyond naive full-context injection to selective, summarized, or hierarchical context provision to maintain performance and cost-effectiveness as content volume increases.
Proprietary Technical Taxonomy
Raw Developer Origin & Technical Request
GitHub Issue
Mar 23, 2026
Repo: Narcooo/inkos
建议优化:长篇写作测试中单章写作时间可高达40分钟
[inkos] 1 post-write errors detected, triggering spot-fix before audit
修复过程,可能高达2000秒以上
spot-fix 需要的上下文所有文件都是全量的
但实际spot-fix 只是修复不是重写
不过这点我有疑问问下大佬的想法
当初是为了方便全部做了全量模板还是还没有开始做特定的优化
这导致一个非常严重的问题
就是写到50章之后修复时间会非常非常长
有些甚至40分钟才能一章 一开始的时候不会这么长时间
Reviser的时间也越来越长
还有在Settler 阶段也是 很多时候都是全量丢给大模型
一开始没啥感觉 到后面我的模型思考炸了@!@
Developer Debate & Comments
不过话又说回来很多很多时候优化过度了会导致小说情节出现问题 不知道大佬设计思路是怎么样的 有时候有些事情鱼和熊掌不可兼得除非大模型上下文又大速度又快 😄
全量注入在早期是很有用的，不过写到长篇导致上下文，记忆和质量的系统性原因，大更新正在加紧测试中！！
原来如此 期待你的大更新 我这边持续测试!
优化卡在phase2这个阶段，因为上传chapter_summaries.md这个文件，这个文件是每一章的概括，越积越大；几十章就能过100kb，传给llm，基本都没回复，或者要等很久；这个要简化下还有机会；
> 全量注入在早期是很有用的，不过写到长篇导致上下文，记忆和质量的系统性原因，大更新正在加紧测试中！！ 兄弟我有个想法，问题就在chapter_summaries.md这个文件越积越多，里面冗余信息太多了；如果说有用的信息，1.hook有关的信息抽出来 2.主要人物状态。。。 其他可以删掉，里面90%以上都是冗余无用信息
Adjacent Repository Pain Points
Other highly discussed features and pain points extracted from Narcooo/inkos.
Extracted Positioning
Architectural design ideas and questions for an AI novel generation system, focusing on RAG, state management, character intelligence, narrative consistency, and feedback loops
Advanced architectural design for scalable, consistent, and intelligent AI novel generation, addressing complex narrative challenges
Top Replies
有一说一，都在做这个啊_(:з」∠)_我也在折腾，不过是基于AI小镇那套思维，给主角和NPC一些自主能动性，自主动的按照自己人设运行，但是运行过程又全程基于游戏，回头再看游戏日志就好，不过现在游戏层卡着我有点...
> 有一说一，都在做这个啊_(:з」∠)_我也在折腾，不过是基于AI小镇那套思维，给主角和NPC一些自主能动性，自主动的按照自己人设运行，但是运行过程又全程基于游戏，回头再看游戏日志就好，不过现在游戏层卡着我有...
我觉得你设计的挺好的，也欢迎pr，有一点就是我个人不太推荐RAG。
Extracted Positioning
Inconsistent API key validation between `inkos doctor` and `inkos write next`, leading to 401 errors during chapter generation
Consistent and reliable API key validation across all operational modes
Top Replies
nkos config show-global 能看到配置如下： ``` (base) jayrome@MacBookPro my-xhnovel % inkos config show-global # InkOS Global LLM Configuration INKOS_LLM_PROVIDER=openai INKOS_LLM_BASE_URL=https://da...
nkos config show-global 能看到配置如下： ``` (base) jayrome@MacBookPro my-xhnovel % inkos config show-global # InkOS Global LLM Configuration INKOS_LLM_PROVIDER=openai INKOS_LLM_BASE_URL=https://da...
检查一下API_KEY是否正确，一般就是key有问题，你可以让豆包或者deepseek帮你写个测试脚本，测试一下
Extracted Positioning
Chapter generation stalling or 'breaking' mid-process, particularly for new books and the first chapter
Reliable and complete chapter generation for new projects
Top Replies
hello，你好，请问哪里获取对呀claude 的 api key呢，有什么渠道购买嘛
> hello，你好，请问哪里获取对呀claude 的 api key呢，有什么渠道购买嘛 这里可以买，还便宜：https://ai-api.db-kj.com/register?aff=uGqz
> hello，你好，请问哪里获取对呀claude 的 api key呢，有什么渠道购买嘛 实测发现 GPT-5.4比较好用, 还有项目的超时方面似乎不太完善, 会断流. 我自己改了,目前生成到20章了. 不过长度目前不受控制, 限制的4000...
Extracted Positioning
API key authentication failure when using custom providers and multiple agents/routes
Reliable API key management and authentication for custom LLM providers and multi-agent configurations
Top Replies
我使用inkos doctor，同样的配置是能够通的，两个网址上稍有差别，但是诊断都能通 恢复为全局默认的就正常了
模型缓存 Key 未包含认证信息，共享同一个缓存 key，等作者更新
感谢报告，问题确认，预计将在下一次大规模更新中修复
Extracted Positioning
Data corruption or cascading errors in project files after rewriting specific chapters
Ensuring data consistency and integrity across all project files during content revision
Top Replies
这些问题基本都是函数的discription没写清楚，导致inkos agent调用错了函数造成的； 这个工程的discription和system prompt写的太少了； revise的功能也有问题
感谢报告，问题确认，正在加强描述和提示词，prompt部分和执行层同步会在0.6.3继续调整
哇难得见到原生的作者大大，这个我是从0.5的时候自己做了修改，做了个记忆系统 writenext以及重写的那个接口都做了优化，现在大体实现的一些点完全解决了这个问题。 1.JSON解析，完全重构了JSON解析这一块，更加...
Frequently Asked Questions
Market intelligence mapped to Performance degradation and excessive token usage in long-form content generation due to 'full context injection'.
What is the technical positioning of Performance degradation and excessive token usage in long-form content generation due to 'full context injection'?
Based on our AI analysis of the original developer request, its primary technical positioning is: Optimizing LLM context management for scalability and efficiency in long-form content generation
What is the general sentiment around Performance degradation and excessive token usage in long-form content generation due to 'full context injection'?
Yes, we have tracked 10 direct responses and active debates regarding this specific topic originating from GitHub Issue.
What architecture is tied to Performance degradation and excessive token usage in long-form content generation due to 'full context injection'?
Our proprietary extraction maps Performance degradation and excessive token usage in long-form content generation due to 'full context injection' to adjacent architectural concepts including post-write errors, spot-fix, Reviser, Settler.
Engagement Signals
Cross-Market Term Frequency
Counts stored text matches for foundational terms like post-write errors and spot-fix in analysed public discussions. These counts do not measure adoption.
Stored text matches, not adoption or market-share measurements.
- post-write errors: 1 records
- spot-fix: 1 records
- Reviser: 1 records
- Settler: 1 records
- 全量丢给大模型: 1 records
- 上下文所有文件都是全量的: 1 records
- chapter_summaries.md: 1 records
- hook有关的信息抽出来: 1 records
- 主要人物状态: 1 records
- 冗余无用信息: 1 records
- 大模型上下文又大速度又快: 1 records
- 大更新正在加紧测试中: 1 records