haylan/LLM-Server
1
0 Fork 0
Code Issues 4 Pull Requests Actions Packages Projects Releases Wiki Activity

Research-doc index: current vs. superseded #57

New Issue
Closed
opened 2026-09-24 19:08:18 +00:00 by haylan · 1 comment
haylan commented 2026-09-24 19:08:18 +00:00
Owner
Copy link
Copy source

Part of #55

Question

For each file in docs/research/, classify it as current (still describes the running stack) or superseded by X (LiteLLM-era, replaced choice, removed feature), with a one-line reason. Image-generation docs are marked for deletion. AFK. Output: the index table in the resolution comment.

Part of #55 ## Question For each file in docs/research/, classify it as *current* (still describes the running stack) or *superseded by X* (LiteLLM-era, replaced choice, removed feature), with a one-line reason. Image-generation docs are marked for deletion. AFK. Output: the index table in the resolution comment.
haylan added the wayfinder:task label 2026-09-24 19:08:18 +00:00
haylan added a new dependency 2026-09-24 19:08:33 +00:00
#60 Target doc structure
haylan self-assigned this 2026-09-24 19:09:04 +00:00
haylan commented 2026-09-24 19:10:13 +00:00
Author
Owner
Copy link
Copy source

Resolution: docs/research/ index

Checked each doc against docker-compose.yml (llama-server Qwen3.8-27B, qwen-classifier Qwen3-4B, OmniRoute, lazytainer, qdrant, neo4j). "Linked from" = grep hits outside docs/research/.

File Status Superseded by / reason Linked from
500k-context-2-parallel-agents-gpu-roadmap.md CURRENT Active hardware roadmap for 500k×2 ctx. It already marks docs/server-planing.md as superseded context, so the two docs conflict (see note 3) none
classifier-context-window-scaling.md CURRENT Decision still in force: classifier stays Qwen3-4B at 65536 ctx, no upgrade to 8B/30B-A3B without a second GPU none
colibri-ai-stack-evaluation.md CURRENT Rejected-option record ("watch, don't integrate"). Still valid none
connecting-cloud-providers.md SUPERSEDED by OmniRoute Written for cloud providers behind the LiteLLM gateway. OmniRoute now registers providers through its dashboard none
dashscope-websearch-selfhost-options.md CURRENT Decision still in force: don't fake DashScope. Keep the OmniRoute MCP + SearXNG path none
fast-model-choice.md CURRENT (with drift) Qwen3-4B-Instruct-2507 is still the classifier. Its own addendum records the drift: service is named qwen-classifier, not llama-server-fast, and the quant is Q4_K_XL, not Q8_0. Sizing moved to classifier-context-window-scaling.md docker-compose.yml, docs/coding-cli-setup/qwen-code.md
image-generation-model-choice.md DELETE Image generation (ComfyUI/Qwen-Image) is being removed .env.example, docker-compose.yml (both go away with the removal, #58)
image-generation-options.md DELETE Image generation (ComfyUI) is being removed docker-compose.yml (goes away with the removal, #58)
langchain-pgvector-vs-litellm-pgvector.md SUPERSEDED by OmniRoute migration + closed LiteLLM-knowledgebase map Compared against the litellm-pgvector connector. LiteLLM is gone and no pgvector exists in compose (qdrant/neo4j run instead) none
lazytainer-omniroute-idle-stop.md CURRENT (partly) Still true and still used: lazytainer can't tell OmniRoute health-check pings from real traffic, and the lazytainer labels are still on llama-server. Obsolete part: the switch-model.sh / comfyui GPU-swap guidance, which goes with the removal scripts/switch-model.sh (being deleted)
litellm-knowledgebase.md SUPERSEDED by OmniRoute migration Covers LiteLLM's vector-store feature. LiteLLM was removed, and the knowledgebase map was closed none
litellm-searxng-search.md SUPERSEDED by omniroute-qwen-websearch.md Search now runs as SearXNG registered in OmniRoute, not through LiteLLM's web-search. Still cited for the search.home DNS/host-IP gotcha, so the pointers need retargeting README.md, .env.example, scripts/update.sh
omniroute-account-semaphore-timeout.md CURRENT Timeout #1 of 3: OmniRoute's hardcoded 30s per-connection semaphore docs/coding-cli-setup/index.md
omniroute-direct-response-timeout-outage-2026-09-15.md CURRENT Timeout #3 of 3 (OMNIROUTE_DIRECT_HEADERS_TIMEOUT_MS) plus the ROCm stall incident. That key is not set in compose/.env.example today none
omniroute-memory-tool-injection-breaks-classifier.md CURRENT OmniRoute memory tools must stay off for the classifier. Still applies none
omniroute-non-ping-sse-stream-timeout.md CURRENT Timeout #2 of 3. Explains the --cache-reuse/LLAMA_PARALLEL and STREAM_IDLE_TIMEOUT_MS settings in compose .env.example, docker-compose.yml
omniroute-qwen-websearch.md CURRENT Qwen Code → OmniRoute routing + MCP/SearXNG web search, as live none
opencode-auto-compact.md CURRENT Covers OpenCode client-side compaction. Independent of the gateway none
opencode-cli-setup.md CURRENT (with drift) Provider JSON shape is still valid, but its baseURL example points straight at llama.cpp :8080. The live setup goes through OmniRoute, per docs/coding-cli-setup/opencode.md docs/coding-cli-setup/opencode.md
ponytail-audit-2026-09-09.md SUPERSEDED (point-in-time) Snapshot audit. Finding 1 (downloader-fast) was fixed in 78adbf4. The rest are either done or folded into #56/#61 none
proxy-shadow-pricing.md SUPERSEDED by OmniRoute migration Price is wired into LiteLLM config.yaml model_info, which no longer exists. The Sonnet reference-price choice is only history now; nothing implements shadow cost none
proxy-tool-choice.md SUPERSEDED by OmniRoute Picked LiteLLM proxy. Replaced by OmniRoute (migration complete) docs/proxy-request-priority.md, docs/proxy-key-onboarding.md
qwen3.8-27b-quant.md CURRENT UD-Q4_K_XL is still the default LLAMA_MODEL_FILE in compose .env.example
qwen3.8-27b-tool-calling.md CURRENT Qwen3.8 tool-call parser caveat for the model still served (--jinja). Clients now reach the model through OmniRoute's /v1/messages, not llama.cpp's shim, but the parser finding still applies README.md, docs/coding-cli-setup/index.md
rocm-gpu-pin-and-render-group.md CURRENT (name drift) The GPU_MAX_HW_QUEUES=1 + numeric group_add GIDs fixes are live on llama-server and qwen-classifier. The doc says llama-server-fast, which is now qwen-classifier docker-compose.yml, .env.example, scripts/update.sh
voidllm-evaluation.md SUPERSEDED by OmniRoute Compared VoidLLM against LiteLLM. Both were moot once OmniRoute was chosen none

Totals: 15 CURRENT (4 with drift noted), 9 SUPERSEDED, 2 DELETE.

Notes for execution:

  1. Non-research files that link to SUPERSEDED docs need a new pointer or a superseded annotation: README.md, .env.example and scripts/update.sh → litellm-searxng-search.md; docs/proxy-request-priority.md and docs/proxy-key-onboarding.md → proxy-tool-choice.md.
  2. Links to the two DELETE docs and to lazytainer-omniroute-idle-stop.md from switch-model.sh disappear with the #58 removal. Check that none survive.
  3. Conflict: 500k-roadmap declares docs/server-planing.md superseded, but #55 says keep server-planing.md. The index should say which build plan is active. Flag for #56/#60.
  4. The docs stay verbatim per #55. The index carries the status, and "drift" items get a one-line note in the index, not an edit to the doc.
## Resolution: docs/research/ index Checked each doc against `docker-compose.yml` (llama-server Qwen3.8-27B, qwen-classifier Qwen3-4B, OmniRoute, lazytainer, qdrant, neo4j). "Linked from" = grep hits outside `docs/research/`. | File | Status | Superseded by / reason | Linked from | |---|---|---|---| | 500k-context-2-parallel-agents-gpu-roadmap.md | CURRENT | Active hardware roadmap for 500k×2 ctx. It already marks `docs/server-planing.md` as superseded context, so the two docs conflict (see note 3) | none | | classifier-context-window-scaling.md | CURRENT | Decision still in force: classifier stays Qwen3-4B at 65536 ctx, no upgrade to 8B/30B-A3B without a second GPU | none | | colibri-ai-stack-evaluation.md | CURRENT | Rejected-option record ("watch, don't integrate"). Still valid | none | | connecting-cloud-providers.md | SUPERSEDED by OmniRoute | Written for cloud providers behind the LiteLLM gateway. OmniRoute now registers providers through its dashboard | none | | dashscope-websearch-selfhost-options.md | CURRENT | Decision still in force: don't fake DashScope. Keep the OmniRoute MCP + SearXNG path | none | | fast-model-choice.md | CURRENT (with drift) | Qwen3-4B-Instruct-2507 is still the classifier. Its own addendum records the drift: service is named `qwen-classifier`, not `llama-server-fast`, and the quant is Q4_K_XL, not Q8_0. Sizing moved to classifier-context-window-scaling.md | docker-compose.yml, docs/coding-cli-setup/qwen-code.md | | image-generation-model-choice.md | DELETE | Image generation (ComfyUI/Qwen-Image) is being removed | .env.example, docker-compose.yml (both go away with the removal, #58) | | image-generation-options.md | DELETE | Image generation (ComfyUI) is being removed | docker-compose.yml (goes away with the removal, #58) | | langchain-pgvector-vs-litellm-pgvector.md | SUPERSEDED by OmniRoute migration + closed LiteLLM-knowledgebase map | Compared against the litellm-pgvector connector. LiteLLM is gone and no pgvector exists in compose (qdrant/neo4j run instead) | none | | lazytainer-omniroute-idle-stop.md | CURRENT (partly) | Still true and still used: lazytainer can't tell OmniRoute health-check pings from real traffic, and the lazytainer labels are still on llama-server. Obsolete part: the switch-model.sh / comfyui GPU-swap guidance, which goes with the removal | scripts/switch-model.sh (being deleted) | | litellm-knowledgebase.md | SUPERSEDED by OmniRoute migration | Covers LiteLLM's vector-store feature. LiteLLM was removed, and the knowledgebase map was closed | none | | litellm-searxng-search.md | SUPERSEDED by omniroute-qwen-websearch.md | Search now runs as SearXNG registered in OmniRoute, not through LiteLLM's web-search. Still cited for the search.home DNS/host-IP gotcha, so the pointers need retargeting | README.md, .env.example, scripts/update.sh | | omniroute-account-semaphore-timeout.md | CURRENT | Timeout #1 of 3: OmniRoute's hardcoded 30s per-connection semaphore | docs/coding-cli-setup/index.md | | omniroute-direct-response-timeout-outage-2026-09-15.md | CURRENT | Timeout #3 of 3 (`OMNIROUTE_DIRECT_HEADERS_TIMEOUT_MS`) plus the ROCm stall incident. That key is not set in compose/.env.example today | none | | omniroute-memory-tool-injection-breaks-classifier.md | CURRENT | OmniRoute memory tools must stay off for the classifier. Still applies | none | | omniroute-non-ping-sse-stream-timeout.md | CURRENT | Timeout #2 of 3. Explains the `--cache-reuse`/`LLAMA_PARALLEL` and `STREAM_IDLE_TIMEOUT_MS` settings in compose | .env.example, docker-compose.yml | | omniroute-qwen-websearch.md | CURRENT | Qwen Code → OmniRoute routing + MCP/SearXNG web search, as live | none | | opencode-auto-compact.md | CURRENT | Covers OpenCode client-side compaction. Independent of the gateway | none | | opencode-cli-setup.md | CURRENT (with drift) | Provider JSON shape is still valid, but its baseURL example points straight at llama.cpp `:8080`. The live setup goes through OmniRoute, per docs/coding-cli-setup/opencode.md | docs/coding-cli-setup/opencode.md | | ponytail-audit-2026-09-09.md | SUPERSEDED (point-in-time) | Snapshot audit. Finding 1 (`downloader-fast`) was fixed in 78adbf4. The rest are either done or folded into #56/#61 | none | | proxy-shadow-pricing.md | SUPERSEDED by OmniRoute migration | Price is wired into LiteLLM `config.yaml` `model_info`, which no longer exists. The Sonnet reference-price choice is only history now; nothing implements shadow cost | none | | proxy-tool-choice.md | SUPERSEDED by OmniRoute | Picked LiteLLM proxy. Replaced by OmniRoute (migration complete) | docs/proxy-request-priority.md, docs/proxy-key-onboarding.md | | qwen3.8-27b-quant.md | CURRENT | UD-Q4_K_XL is still the default `LLAMA_MODEL_FILE` in compose | .env.example | | qwen3.8-27b-tool-calling.md | CURRENT | Qwen3.8 tool-call parser caveat for the model still served (`--jinja`). Clients now reach the model through OmniRoute's `/v1/messages`, not llama.cpp's shim, but the parser finding still applies | README.md, docs/coding-cli-setup/index.md | | rocm-gpu-pin-and-render-group.md | CURRENT (name drift) | The `GPU_MAX_HW_QUEUES=1` + numeric `group_add` GIDs fixes are live on llama-server and qwen-classifier. The doc says `llama-server-fast`, which is now `qwen-classifier` | docker-compose.yml, .env.example, scripts/update.sh | | voidllm-evaluation.md | SUPERSEDED by OmniRoute | Compared VoidLLM against LiteLLM. Both were moot once OmniRoute was chosen | none | **Totals:** 15 CURRENT (4 with drift noted), 9 SUPERSEDED, 2 DELETE. **Notes for execution:** 1. Non-research files that link to SUPERSEDED docs need a new pointer or a superseded annotation: README.md, .env.example and scripts/update.sh → litellm-searxng-search.md; docs/proxy-request-priority.md and docs/proxy-key-onboarding.md → proxy-tool-choice.md. 2. Links to the two DELETE docs and to lazytainer-omniroute-idle-stop.md from switch-model.sh disappear with the #58 removal. Check that none survive. 3. Conflict: 500k-roadmap declares `docs/server-planing.md` superseded, but #55 says keep server-planing.md. The index should say which build plan is active. Flag for #56/#60. 4. The docs stay verbatim per #55. The index carries the status, and "drift" items get a one-line note in the index, not an edit to the doc.
Sign in to join this conversation.
No Branch/Tag Specified
Labels
Clear labels
wayfinder:grilling
Wayfinder grilling ticket
 wayfinder:map
Wayfinder map issue
 wayfinder:prototype
Wayfinder prototype ticket
 wayfinder:research
Wayfinder research ticket
 wayfinder:task
Wayfinder task ticket
No labels wayfinder:task
Milestone
No items
No Milestone
Projects
Clear projects
No projects
Assignees
Clear assignees
haylan (Arthur Erlich)
No Assignees haylan
1 Participants
Notifications
Due Date
No due date set.
Blocks
#60 Target doc structure
haylan/LLM-Server
Reference: haylan/LLM-Server#57