deepdoc 1.2.0__tar.gz → 1.3.0__tar.gz

{deepdoc-1.2.0 → deepdoc-1.3.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: deepdoc
-Version: 1.2.0
+Version: 1.3.0
 Summary: Auto-generate beautiful docs from any codebase
 Author: Pranav Kumar
 License: MIT
@@ -310,6 +310,22 @@ deepdoc config set output_dir documentation            # Change output dir
 deepdoc config set llm.api_key_env AZURE_API_KEY       # Change API key env var
 ```
 
+### `deepdoc benchmark`
+
+Run planner benchmark cases and optionally generate a combined docs+chatbot quality scorecard.
+
+```bash
+deepdoc benchmark --catalog benchmarks/catalog.json
+deepdoc benchmark --repo /path/to/repo --gold benchmarks/gold.json
+deepdoc benchmark --catalog benchmarks/catalog.json --chatbot-eval benchmarks/chatbot_eval.json
+deepdoc benchmark --catalog benchmarks/catalog.json --chatbot-eval benchmarks/chatbot_eval.json --scorecard-out .deepdoc/quality_scorecard.json --strict-scorecard
+deepdoc benchmark --generated-root /Users/apple/autodoc/docs --scorecard-out /Users/apple/autodoc/docs/_scorecards/latest.json
+```
+
+Use `--strict-scorecard` to fail the command when completeness gates are not met.
+
+When you do not have a hand-written benchmark catalog or chatbot eval file yet, use artifact mode (`--generated-root` or `--artifact-repo`) to compute a provisional scorecard directly from persisted `.deepdoc/` outputs.
+
 ---
 
 ## LLM Provider Setup
@@ -618,7 +634,9 @@ chatbot:
     base_url: ""
     api_version: ""
     temperature: 0.1
-    max_tokens: 16000
+    max_tokens: 24000
+    continuation_retries: 2                   # Auto-continue if answer ends abruptly
+    continuation_context_chars: 12000         # Tail chars included in continuation prompt
 
   embeddings:                                 # LLM used for embedding code/docs
     provider: "azure"
@@ -643,16 +661,20 @@ chatbot:
     top_k_code: 15
     top_k_artifact: 8
     top_k_docs: 6
-    top_k_relationship: 6
+    top_k_relationship: 8
     candidate_top_k_code: 30
     candidate_top_k_artifact: 16
     candidate_top_k_docs: 12
     candidate_top_k_relationship: 12
     max_prompt_code_chunks: 12
     max_prompt_artifact_chunks: 6
-    max_prompt_doc_chunks: 4
-    max_prompt_relationship_chunks: 4
-    max_prompt_chars: 200000
+    max_prompt_doc_chunks: 6
+    max_prompt_relationship_chunks: 6
+    max_prompt_chars: 120000
+    fast_mode_use_llm_retrieval_steps: false  # Fast mode skips expansion/rerank by default
+    fast_mode_iterative_retrieval: false      # Fast mode skips second-pass follow-up retrieval
+    fast_mode_max_prompt_chars: 90000         # Smaller prompt budget for faster /query answers
+    deep_mode_max_prompt_chars: 140000        # Larger budget for /deep-research synthesis
     lexical_retrieval: true
     lexical_candidate_limit: 24
     query_expansion: true
@@ -666,7 +688,8 @@ chatbot:
     graph_neighbor_relationship_chunks_per_file: 2
     graph_neighbor_max_docs: 4
     rerank: true
-    rerank_candidate_limit: 20
+    rerank_candidate_limit: 32
+    rerank_candidate_limit_per_kind: 8
     rerank_preview_chars: 450
     stitch_adjacent_code_chunks: true
     stitch_max_adjacent_chunks: 2
@@ -674,7 +697,8 @@ chatbot:
     live_fallback_max_files: 6
     live_fallback_max_per_file: 2
     live_fallback_context_lines: 12
-    deep_research_chunk_chars: 1600
+    deep_research_chunk_chars: 3200
+    deep_research_top_k: 10
 
   chunking:
     code_chunk_lines: 120
@@ -709,7 +733,9 @@ chatbot:
 | `chatbot.answer.base_url` | `""` | Custom endpoint (for Azure, Ollama, etc.) |
 | `chatbot.answer.api_version` | `""` | Azure API version string |
 | `chatbot.answer.temperature` | `0.1` | Sampling temperature (lower = more deterministic) |
-| `chatbot.answer.max_tokens` | `16000` | Max tokens per answer |
+| `chatbot.answer.max_tokens` | `24000` | Max tokens per answer |
+| `chatbot.answer.continuation_retries` | `2` | Extra completion attempts when an answer appears truncated |
+| `chatbot.answer.continuation_context_chars` | `12000` | Number of trailing chars passed when asking the model to continue |
 | **Embeddings LLM** | | |
 | `chatbot.embeddings.provider` | `azure` | Provider for the embedding model |
 | `chatbot.embeddings.model` | `azure/text-embedding-3-large` | Embedding model |
@@ -721,16 +747,20 @@ chatbot:
 | `chatbot.retrieval.top_k_code` | `15` | Top code chunks retrieved per query |
 | `chatbot.retrieval.top_k_artifact` | `8` | Top artifact chunks retrieved per query |
 | `chatbot.retrieval.top_k_docs` | `6` | Top generated-doc and repo-doc chunks retrieved per query |
-| `chatbot.retrieval.top_k_relationship` | `6` | Top relationship chunks retrieved per query |
+| `chatbot.retrieval.top_k_relationship` | `8` | Top relationship chunks retrieved per query |
 | `chatbot.retrieval.candidate_top_k_code` | `30` | Candidate code chunks gathered before reranking |
 | `chatbot.retrieval.candidate_top_k_artifact` | `16` | Candidate artifact chunks gathered before reranking |
 | `chatbot.retrieval.candidate_top_k_docs` | `12` | Candidate doc chunks gathered before reranking |
 | `chatbot.retrieval.candidate_top_k_relationship` | `12` | Candidate relationship chunks gathered before reranking |
 | `chatbot.retrieval.max_prompt_code_chunks` | `12` | Max code chunks included in the final prompt |
 | `chatbot.retrieval.max_prompt_artifact_chunks` | `6` | Max artifact chunks in the final prompt |
-| `chatbot.retrieval.max_prompt_doc_chunks` | `4` | Max doc chunks in the final prompt |
-| `chatbot.retrieval.max_prompt_relationship_chunks` | `4` | Max relationship chunks included in the final prompt |
-| `chatbot.retrieval.max_prompt_chars` | `200000` | Total character budget for the assembled prompt |
+| `chatbot.retrieval.max_prompt_doc_chunks` | `6` | Max doc chunks in the final prompt |
+| `chatbot.retrieval.max_prompt_relationship_chunks` | `6` | Max relationship chunks included in the final prompt |
+| `chatbot.retrieval.max_prompt_chars` | `120000` | Default character budget for assembled prompts |
+| `chatbot.retrieval.fast_mode_use_llm_retrieval_steps` | `false` | In `/query` fast mode, disable LLM query expansion and reranking |
+| `chatbot.retrieval.fast_mode_iterative_retrieval` | `false` | In `/query` fast mode, disable iterative follow-up retrieval |
+| `chatbot.retrieval.fast_mode_max_prompt_chars` | `90000` | Prompt budget used by `/query` fast mode |
+| `chatbot.retrieval.deep_mode_max_prompt_chars` | `140000` | Prompt budget used by `/deep-research` |
 | `chatbot.retrieval.lexical_retrieval` | `true` | Blend exact-match retrieval with embedding retrieval |
 | `chatbot.retrieval.lexical_candidate_limit` | `24` | Max lexical candidates gathered before merge/rerank |
 | `chatbot.retrieval.query_expansion` | `true` | Use LLM to generate alternative search queries |
@@ -744,7 +774,8 @@ chatbot:
 | `chatbot.retrieval.graph_neighbor_relationship_chunks_per_file` | `2` | Relationship chunks per linked file during graph expansion |
 | `chatbot.retrieval.graph_neighbor_max_docs` | `4` | Max linked docs pulled in during graph expansion |
 | `chatbot.retrieval.rerank` | `true` | Use LLM to rerank retrieved chunks |
-| `chatbot.retrieval.rerank_candidate_limit` | `20` | Max candidates sent to the reranker |
+| `chatbot.retrieval.rerank_candidate_limit` | `32` | Max candidates sent to the reranker |
+| `chatbot.retrieval.rerank_candidate_limit_per_kind` | `8` | Per-kind candidate cap before filling the global rerank pool |
 | `chatbot.retrieval.rerank_preview_chars` | `450` | Characters of each chunk shown to the reranker |
 | `chatbot.retrieval.stitch_adjacent_code_chunks` | `true` | Expand exact-match code hits with adjacent windows from the same file |
 | `chatbot.retrieval.stitch_max_adjacent_chunks` | `2` | Max adjacent code windows stitched onto a top hit |
@@ -752,7 +783,8 @@ chatbot:
 | `chatbot.retrieval.live_fallback_max_files` | `6` | Max repo files inspected during a deep-research live fallback |
 | `chatbot.retrieval.live_fallback_max_per_file` | `2` | Max fallback snippets returned per inspected file |
 | `chatbot.retrieval.live_fallback_context_lines` | `12` | Lines per fallback snippet around each exact match |
-| `chatbot.retrieval.deep_research_chunk_chars` | `1600` | Max chars per evidence chunk passed into deep-research step answers |
+| `chatbot.retrieval.deep_research_chunk_chars` | `3200` | Max chars per evidence chunk passed into deep-research step answers |
+| `chatbot.retrieval.deep_research_top_k` | `10` | Retrieved chunks per deep-research sub-question |
 | **Chunking** | | |
 | `chatbot.chunking.code_chunk_lines` | `120` | Lines per code chunk |
 | `chatbot.chunking.code_chunk_overlap` | `20` | Overlap lines between code chunks |
@@ -829,22 +861,24 @@ During `deepdoc generate`, six corpora are built and stored in `.deepdoc/chatbot
 
 ### Chatbot Query Pipeline
 
-When a user asks a question, the backend runs a multi-step retrieval pipeline:
+When a user asks a question, the backend runs a mode-aware retrieval pipeline:
 
-1. **Query expansion** — The LLM generates up to 3 alternative search queries to improve recall.
+1. **Query expansion** — In default/deep mode, the LLM can generate alternative search queries to improve recall. Fast mode disables this by default.
 2. **Embedding** — All queries are embedded using the configured embedding model.
 3. **Hybrid retrieval** — FAISS similarity search and exact-match lexical search both gather candidates from each corpus.
-4. **Follow-up retrieval** — The backend can derive focused second-pass searches and pull linked files/docs via graph-neighbor expansion.
+4. **Follow-up retrieval** — The backend can derive focused second-pass searches and pull linked files/docs via graph-neighbor expansion. Fast mode can skip follow-up queries for lower latency.
 5. **Chunk stitching** — Exact-match code hits can pull adjacent code windows from the same file so larger implementations survive chunk boundaries.
-6. **Reranking** — The LLM scores and reranks the retrieved chunks for relevance.
+6. **Reranking** — In default/deep mode, the LLM can rerank candidates for relevance. Fast mode disables this by default.
 7. **Prompt assembly** — Query-type-aware budgets reserve space for the most important evidence types within the character budget.
-8. **Answer generation** — The answer LLM produces a grounded response with code, artifact, doc, repo-doc, relationship, and live-fallback citations when used.
+8. **Answer generation + continuity guard** — The answer LLM produces a grounded response, and if the output appears truncated (for example ending on a dangling heading), DeepDoc retries with a continuation prompt so the response finishes cleanly.
 
 `POST /deep-research` uses the same indexed corpora first, but it can also inspect a small bounded set of live repo files when exact-match evidence is missing from the index. This fallback respects the repo's exclude rules, skips oversized/binary files, and is only used in deep research mode.
 
+`POST /query` and `POST /deep-research` now return `response_mode` in the payload (`fast`, `deep`, or `default`) so clients can confirm which retrieval profile generated the result.
+
 ### Chatbot API Endpoints
 
-The generated `chatbot_backend/` exposes two endpoints:
+The generated `chatbot_backend/` exposes three endpoints:
 
 **Health check:**
 ```
@@ -865,6 +899,19 @@ POST /query
 
 The response includes the answer text, code citations (file path + line range), artifact citations, and links to relevant generated doc pages.
 
+`/query` is optimized for speed: it runs retrieval in fast mode (no LLM query expansion/rerank by default) and returns an answer plus citations.
+
+**Retrieve context only (no answer generation):**
+```
+POST /query-context
+{
+  "question": "Where is reshipping implemented?",
+  "history": []
+}
+```
+
+`/query-context` returns selected citations/chunks only. Use this endpoint to inspect retrieval quality independently from answer generation.
+
 ### Deploying the Chatbot
 
 For local development, `deepdoc serve` handles everything automatically. For production:

{deepdoc-1.2.0 → deepdoc-1.3.0}/README.md

@@ -271,6 +271,22 @@ deepdoc config set output_dir documentation            # Change output dir
 deepdoc config set llm.api_key_env AZURE_API_KEY       # Change API key env var
 ```
 
+### `deepdoc benchmark`
+
+Run planner benchmark cases and optionally generate a combined docs+chatbot quality scorecard.
+
+```bash
+deepdoc benchmark --catalog benchmarks/catalog.json
+deepdoc benchmark --repo /path/to/repo --gold benchmarks/gold.json
+deepdoc benchmark --catalog benchmarks/catalog.json --chatbot-eval benchmarks/chatbot_eval.json
+deepdoc benchmark --catalog benchmarks/catalog.json --chatbot-eval benchmarks/chatbot_eval.json --scorecard-out .deepdoc/quality_scorecard.json --strict-scorecard
+deepdoc benchmark --generated-root /Users/apple/autodoc/docs --scorecard-out /Users/apple/autodoc/docs/_scorecards/latest.json
+```
+
+Use `--strict-scorecard` to fail the command when completeness gates are not met.
+
+When you do not have a hand-written benchmark catalog or chatbot eval file yet, use artifact mode (`--generated-root` or `--artifact-repo`) to compute a provisional scorecard directly from persisted `.deepdoc/` outputs.
+
 ---
 
 ## LLM Provider Setup
@@ -579,7 +595,9 @@ chatbot:
     base_url: ""
     api_version: ""
     temperature: 0.1
-    max_tokens: 16000
+    max_tokens: 24000
+    continuation_retries: 2                   # Auto-continue if answer ends abruptly
+    continuation_context_chars: 12000         # Tail chars included in continuation prompt
 
   embeddings:                                 # LLM used for embedding code/docs
     provider: "azure"
@@ -604,16 +622,20 @@ chatbot:
     top_k_code: 15
     top_k_artifact: 8
     top_k_docs: 6
-    top_k_relationship: 6
+    top_k_relationship: 8
     candidate_top_k_code: 30
     candidate_top_k_artifact: 16
     candidate_top_k_docs: 12
     candidate_top_k_relationship: 12
     max_prompt_code_chunks: 12
     max_prompt_artifact_chunks: 6
-    max_prompt_doc_chunks: 4
-    max_prompt_relationship_chunks: 4
-    max_prompt_chars: 200000
+    max_prompt_doc_chunks: 6
+    max_prompt_relationship_chunks: 6
+    max_prompt_chars: 120000
+    fast_mode_use_llm_retrieval_steps: false  # Fast mode skips expansion/rerank by default
+    fast_mode_iterative_retrieval: false      # Fast mode skips second-pass follow-up retrieval
+    fast_mode_max_prompt_chars: 90000         # Smaller prompt budget for faster /query answers
+    deep_mode_max_prompt_chars: 140000        # Larger budget for /deep-research synthesis
     lexical_retrieval: true
     lexical_candidate_limit: 24
     query_expansion: true
@@ -627,7 +649,8 @@ chatbot:
     graph_neighbor_relationship_chunks_per_file: 2
     graph_neighbor_max_docs: 4
     rerank: true
-    rerank_candidate_limit: 20
+    rerank_candidate_limit: 32
+    rerank_candidate_limit_per_kind: 8
     rerank_preview_chars: 450
     stitch_adjacent_code_chunks: true
     stitch_max_adjacent_chunks: 2
@@ -635,7 +658,8 @@ chatbot:
     live_fallback_max_files: 6
     live_fallback_max_per_file: 2
     live_fallback_context_lines: 12
-    deep_research_chunk_chars: 1600
+    deep_research_chunk_chars: 3200
+    deep_research_top_k: 10
 
   chunking:
     code_chunk_lines: 120
@@ -670,7 +694,9 @@ chatbot:
 | `chatbot.answer.base_url` | `""` | Custom endpoint (for Azure, Ollama, etc.) |
 | `chatbot.answer.api_version` | `""` | Azure API version string |
 | `chatbot.answer.temperature` | `0.1` | Sampling temperature (lower = more deterministic) |
-| `chatbot.answer.max_tokens` | `16000` | Max tokens per answer |
+| `chatbot.answer.max_tokens` | `24000` | Max tokens per answer |
+| `chatbot.answer.continuation_retries` | `2` | Extra completion attempts when an answer appears truncated |
+| `chatbot.answer.continuation_context_chars` | `12000` | Number of trailing chars passed when asking the model to continue |
 | **Embeddings LLM** | | |
 | `chatbot.embeddings.provider` | `azure` | Provider for the embedding model |
 | `chatbot.embeddings.model` | `azure/text-embedding-3-large` | Embedding model |
@@ -682,16 +708,20 @@ chatbot:
 | `chatbot.retrieval.top_k_code` | `15` | Top code chunks retrieved per query |
 | `chatbot.retrieval.top_k_artifact` | `8` | Top artifact chunks retrieved per query |
 | `chatbot.retrieval.top_k_docs` | `6` | Top generated-doc and repo-doc chunks retrieved per query |
-| `chatbot.retrieval.top_k_relationship` | `6` | Top relationship chunks retrieved per query |
+| `chatbot.retrieval.top_k_relationship` | `8` | Top relationship chunks retrieved per query |
 | `chatbot.retrieval.candidate_top_k_code` | `30` | Candidate code chunks gathered before reranking |
 | `chatbot.retrieval.candidate_top_k_artifact` | `16` | Candidate artifact chunks gathered before reranking |
 | `chatbot.retrieval.candidate_top_k_docs` | `12` | Candidate doc chunks gathered before reranking |
 | `chatbot.retrieval.candidate_top_k_relationship` | `12` | Candidate relationship chunks gathered before reranking |
 | `chatbot.retrieval.max_prompt_code_chunks` | `12` | Max code chunks included in the final prompt |
 | `chatbot.retrieval.max_prompt_artifact_chunks` | `6` | Max artifact chunks in the final prompt |
-| `chatbot.retrieval.max_prompt_doc_chunks` | `4` | Max doc chunks in the final prompt |
-| `chatbot.retrieval.max_prompt_relationship_chunks` | `4` | Max relationship chunks included in the final prompt |
-| `chatbot.retrieval.max_prompt_chars` | `200000` | Total character budget for the assembled prompt |
+| `chatbot.retrieval.max_prompt_doc_chunks` | `6` | Max doc chunks in the final prompt |
+| `chatbot.retrieval.max_prompt_relationship_chunks` | `6` | Max relationship chunks included in the final prompt |
+| `chatbot.retrieval.max_prompt_chars` | `120000` | Default character budget for assembled prompts |
+| `chatbot.retrieval.fast_mode_use_llm_retrieval_steps` | `false` | In `/query` fast mode, disable LLM query expansion and reranking |
+| `chatbot.retrieval.fast_mode_iterative_retrieval` | `false` | In `/query` fast mode, disable iterative follow-up retrieval |
+| `chatbot.retrieval.fast_mode_max_prompt_chars` | `90000` | Prompt budget used by `/query` fast mode |
+| `chatbot.retrieval.deep_mode_max_prompt_chars` | `140000` | Prompt budget used by `/deep-research` |
 | `chatbot.retrieval.lexical_retrieval` | `true` | Blend exact-match retrieval with embedding retrieval |
 | `chatbot.retrieval.lexical_candidate_limit` | `24` | Max lexical candidates gathered before merge/rerank |
 | `chatbot.retrieval.query_expansion` | `true` | Use LLM to generate alternative search queries |
@@ -705,7 +735,8 @@ chatbot:
 | `chatbot.retrieval.graph_neighbor_relationship_chunks_per_file` | `2` | Relationship chunks per linked file during graph expansion |
 | `chatbot.retrieval.graph_neighbor_max_docs` | `4` | Max linked docs pulled in during graph expansion |
 | `chatbot.retrieval.rerank` | `true` | Use LLM to rerank retrieved chunks |
-| `chatbot.retrieval.rerank_candidate_limit` | `20` | Max candidates sent to the reranker |
+| `chatbot.retrieval.rerank_candidate_limit` | `32` | Max candidates sent to the reranker |
+| `chatbot.retrieval.rerank_candidate_limit_per_kind` | `8` | Per-kind candidate cap before filling the global rerank pool |
 | `chatbot.retrieval.rerank_preview_chars` | `450` | Characters of each chunk shown to the reranker |
 | `chatbot.retrieval.stitch_adjacent_code_chunks` | `true` | Expand exact-match code hits with adjacent windows from the same file |
 | `chatbot.retrieval.stitch_max_adjacent_chunks` | `2` | Max adjacent code windows stitched onto a top hit |
@@ -713,7 +744,8 @@ chatbot:
 | `chatbot.retrieval.live_fallback_max_files` | `6` | Max repo files inspected during a deep-research live fallback |
 | `chatbot.retrieval.live_fallback_max_per_file` | `2` | Max fallback snippets returned per inspected file |
 | `chatbot.retrieval.live_fallback_context_lines` | `12` | Lines per fallback snippet around each exact match |
-| `chatbot.retrieval.deep_research_chunk_chars` | `1600` | Max chars per evidence chunk passed into deep-research step answers |
+| `chatbot.retrieval.deep_research_chunk_chars` | `3200` | Max chars per evidence chunk passed into deep-research step answers |
+| `chatbot.retrieval.deep_research_top_k` | `10` | Retrieved chunks per deep-research sub-question |
 | **Chunking** | | |
 | `chatbot.chunking.code_chunk_lines` | `120` | Lines per code chunk |
 | `chatbot.chunking.code_chunk_overlap` | `20` | Overlap lines between code chunks |
@@ -790,22 +822,24 @@ During `deepdoc generate`, six corpora are built and stored in `.deepdoc/chatbot
 
 ### Chatbot Query Pipeline
 
-When a user asks a question, the backend runs a multi-step retrieval pipeline:
+When a user asks a question, the backend runs a mode-aware retrieval pipeline:
 
-1. **Query expansion** — The LLM generates up to 3 alternative search queries to improve recall.
+1. **Query expansion** — In default/deep mode, the LLM can generate alternative search queries to improve recall. Fast mode disables this by default.
 2. **Embedding** — All queries are embedded using the configured embedding model.
 3. **Hybrid retrieval** — FAISS similarity search and exact-match lexical search both gather candidates from each corpus.
-4. **Follow-up retrieval** — The backend can derive focused second-pass searches and pull linked files/docs via graph-neighbor expansion.
+4. **Follow-up retrieval** — The backend can derive focused second-pass searches and pull linked files/docs via graph-neighbor expansion. Fast mode can skip follow-up queries for lower latency.
 5. **Chunk stitching** — Exact-match code hits can pull adjacent code windows from the same file so larger implementations survive chunk boundaries.
-6. **Reranking** — The LLM scores and reranks the retrieved chunks for relevance.
+6. **Reranking** — In default/deep mode, the LLM can rerank candidates for relevance. Fast mode disables this by default.
 7. **Prompt assembly** — Query-type-aware budgets reserve space for the most important evidence types within the character budget.
-8. **Answer generation** — The answer LLM produces a grounded response with code, artifact, doc, repo-doc, relationship, and live-fallback citations when used.
+8. **Answer generation + continuity guard** — The answer LLM produces a grounded response, and if the output appears truncated (for example ending on a dangling heading), DeepDoc retries with a continuation prompt so the response finishes cleanly.
 
 `POST /deep-research` uses the same indexed corpora first, but it can also inspect a small bounded set of live repo files when exact-match evidence is missing from the index. This fallback respects the repo's exclude rules, skips oversized/binary files, and is only used in deep research mode.
 
+`POST /query` and `POST /deep-research` now return `response_mode` in the payload (`fast`, `deep`, or `default`) so clients can confirm which retrieval profile generated the result.
+
 ### Chatbot API Endpoints
 
-The generated `chatbot_backend/` exposes two endpoints:
+The generated `chatbot_backend/` exposes three endpoints:
 
 **Health check:**
 ```
@@ -826,6 +860,19 @@ POST /query
 
 The response includes the answer text, code citations (file path + line range), artifact citations, and links to relevant generated doc pages.
 
+`/query` is optimized for speed: it runs retrieval in fast mode (no LLM query expansion/rerank by default) and returns an answer plus citations.
+
+**Retrieve context only (no answer generation):**
+```
+POST /query-context
+{
+  "question": "Where is reshipping implemented?",
+  "history": []
+}
+```
+
+`/query-context` returns selected citations/chunks only. Use this endpoint to inspect retrieval quality independently from answer generation.
+
 ### Deploying the Chatbot
 
 For local development, `deepdoc serve` handles everything automatically. For production:

{deepdoc-1.2.0 → deepdoc-1.3.0}/deepdoc/__init__.py

@@ -1,3 +1,3 @@
 """DeepDoc — Auto-generate beautiful docs from any codebase."""
 
-__version__ = "1.2.0"
+__version__ = "1.3.0"